Ownership consolidation, monopsony power and efficiency:Evidence from the Chinese tobacco industry

Michael Rubens*

Job Market Paper

December 12, 2019Latest version here

Abstract

This paper examines the effects of ownership consolidation among Chinese cigarette

manufacturers on monopsony power, product market power and productive effi-

ciency. I combine a structural model of production and input market competition

with a natural experiment in which firms under certain production quantity thresh-

olds were forced to exit. This consolidation led to an increase in tobacco leaf price

markdowns of 37%, while cigarette price markups fell by 23%. Although the ob-

jective of the consolidation policy was to increase efficiency, I do not find strong

evidence for such gains. Rising monopsony power has important distributional

consequences. The policy of forcing small manufacturers to exit explains 42% of

the increase in income inequality between farmers and manufacturing workers in

the tobacco industry between 2003 and 2006.

Keywords: Market Power, Monopsony, Concentration, Productivity, Inequality

JEL Codes: L10, J42, O25, D33

*KU Leuven and Research Foundation Flanders. E-mail: michael.rubens@kuleuven.be.

I am grateful to Jan De Loecker and Jo Van Biesebroeck for their mentoring and advice. I also thank FrankVerboven, Chad Syverson, Otto Toivanen, Dan Ackerberg, Myrto Kalouptsidi, and Benjamin Vatter for theirhelpful suggestions, and participants at EARIE 2019, IIOC 2019, the Warwick Economics PhD Conferenceand KU Leuven IO and Development workshops. I thank Yale’s Economics department for hosting me asa visiting PhD student. This is the first chapter of my Ph.D. dissertation at KU Leuven. Financial supportfrom the Research Foundation Flanders, Fulbright Program and Belgian American Educational Foundation aregratefully acknowledged. Any remaining errors are solely my responsibility.

1

1 Introduction

There is an ongoing debate about the economy-wide evolution of market concentration (Rossi-Hansberg,Sarte, & Trachter, 2018; Autor, Dorn, Katz, Patterson, & Van Reenen, 2017; Covarrubias, Gutierrez,& Philippon, 2019), and of markups (De Loecker, Eeckhout, & Unger, 2018; De Loecker & Eeck-hout, 2018). In theory, the degree of market concentration can affect both competition and productiveefficiency. Ownership consolidation may induce firms to change their markup of prices over marginalcosts, but they could also adjust the extent to which input prices are marked down below marginalinput products.1 Moreover, production could become more efficient, for instance due to increasingreturns to scale or scale economies. Empirical work on ownership consolidation tends to focus on oneor two of these channels while ruling out the others. In order to fully understand the welfare conse-quences of ownership consolidation, however, all three potential effects need to be jointly examined.

This paper fills this gap by examining the effects of ownership consolidation in the Chinese cigarettemanufacturing industry on both cigarette price markups, tobacco leaf price markdowns, and produc-tive efficiency. Besides the fact that this is a large industry,2 it also provides the ideal setting tostudy consolidation due to quasi-experimental variation in market structure. In 2003, the Chinesegovernment initiated a large consolidation wave during which cigarette manufacturers under specificoutput thresholds were forced to close down. This variation is useful because other sources of marketstructure variation, such as mergers and acquisitions and exit and entry, tend to be endogenous toproductivity and market power.

Another interesting feature of the Chinese tobacco industry is that market concentration increasesalong the value chain. Around 20 million farmers sell leaves to manufacturers, with the numberof manufacturers decreasing from 350 to 150 during the consolidation. In turn, manufacturers sellcigarettes domestically to a monopsonic government-controlled wholesaler.3 We can thus expectmonopsony power to be present along the chain.4

The key driver of monopsony power on leaf markets is the legal obligation of tobacco farmersto sell their entire output at local purchasing stations, where only local manufacturers compete forleaves. Transporting leaves outside these small isolated markets is illegal.5 Such localized markets

1The relationship between concentration, markups and markdowns is theoretically ambiguous, see Syverson (2019).2Annual industry revenue exceeds $7 billion , and 40% of the world’s cigarettes are made in China.3The tobacco industry remained largely domestic even after China’s WTO accession, as exports make up for less than 1%of industry revenue. This eliminates various potentially confounding factors which relate to international trade.

4Public health externalities are, finally, an idiosyncratic aspect of the tobacco industry. I will abstract from these healthconcerns in this paper.

5Counterfeiting and leaf trade on black markets does happen in China. In the absence of data on such illegal transactions,I abstract from illegal trade flows.

2

are widespread across the rural developing world, and are, for instance, a driver of monopsony poweron Indian agricultural markets as well (Chatterjee, 2019). Although farmers can switch crops, this iscostly: an experiment with crop substitution from 2008 found that switching to other crops increasedthe annual revenue per acre of Chinese tobacco farmers by 21% to 110% (Li, Wang, Xia, Tang, &Wang, 2012). The fact that farmers forgo these large revenue gains suggests large crop switchingcosts.6 Leaving agriculture altogether is another possibility, but mostly involves migration. Both theHukou permit system and land tenure uncertainty increase rural-urban migration costs (Minale, 2018).Finally, county governments in tobacco-rich provinces are heavily dependent on fiscal revenue fromthe tobacco industry, and there are reports of tobacco farmers being coerced not to switch towardsother crops by local authorities (Wang, 2013).

I start by providing reduced-form evidence for the effects of the consolidation on both leaf andcigarette prices. I compare manufacturers which competed with firms that produced below the exitthreshold at the onset of the consolidation in 2003 (the treatment group) with manufacturers withoutsuch competitors (the control group). I find that for firms in the treatment group, leaf prices fell by40% to 64% more compared to firms in the control group between 2003 and 2006. Labor wagesdid, in contrast, fall by 6% to 15%, but this drop was not statistically significant. Finally, factory-gate cigarette prices fell between 24-34% more for the treatment group. Although this reduced-formevidence is arguably causal, it is not sufficient to draw conclusions about the underlying mechanism.Prices could have changed due to changes in market power on product or input markets, but alsodue to changes in productive efficiency. In order to identify the exact mechanisms through which theconsolidation affected prices, more structure is needed.

I therefore construct and estimate a structural model in which cigarette price markups, leaf pricemarkdowns and manufacturing efficiency are jointly identified. I combine firm-level data on costs,prices and production quantities and build on the cost-side approach of Hall (1986) and De Loeckerand Warzynski (2012), with the key distinction that I allow for a subset of inputs to be non-substitutable.7

Substitutability between tobacco leaf and manufacturing workers is low, so I use a Leontief model inleaf and labor.8 I combine this approach with a model of leaf supply.9 I start by estimating the produc-tion function for cigarettes, and discuss the conditions under which prior control function approachesidentify the production function when input prices are endogenous. Next, I use the productivity resid-

6These switching costs are a key feature of the agricultural economics literature (Song, Zhao, & Swinton, 2011; Scott,2013).

7In their analysis of market power in the beer industry, De Loecker and Scott (2016) also allowed for a non-substitutableinput, but not for input market power.

8I estimate the elasticity of substitution between leaf and labor and find that a Leontief cannot be rejected, while a Cobb-Douglas production function can be rejected.

9More structure on input market competition was also placed in Tortarolo and Zarate (2018), but with a very differentidentification strategy, and while still relying on substitutability between all inputs.

3

uals as input demand shifters to identify the tobacco leaf supply function and infer markdowns. Thecondition for this exclusion restriction to hold is that, conditional on the leaf price, manufacturingproductivity shocks are excluded from farmer utility. Finally, I bring the production model and in-put supply model together to infer markups without imposing any additional assumptions on howmanufacturers compete on cigarette markets.

I find that markups of cigarette producers were not significantly different from one, meaning thatprices were equal to marginal costs. Leaf price markdowns were, in contrast, large: the averagecigarette manufacturer paid its tobacco farmers 38% of their marginal revenue product. This is muchless compared to most prior monopsony studies such as Goolsbee and Syverson (2019) for US tenure-track college professors (83%) and Oaxaca and Ransom (2010) for US grocery clerks (70%). Thecombination of low markups and high markdowns of manufacturers is consistent with the fact thatthey bought from many small farmers, but sold to a single large buyer.10

Having estimated both markups, markdowns and productive efficiency, I examine how the consol-idation policy affected these three outcomes. I find that markdowns increased on average by 37%more in the treatment group relative to the control group when defining input markets at the province-level, and by even more when defining input markets more narrowly. The markdown increase was thelargest in markets with high numbers of immigrants, non-registered inhabitants and high unemploy-ment rates. Markups of manufacturers fell, in contrast, by 23% as the wholesaler used its own buyingpower to decrease factory-gate prices.

Finally, I examine the effects of the consolidation on redistribution and scale economies. By in-creasing markdowns on tobacco leaf markets, but not on manufacturing labor markets, the consoli-dation led to an increase of 42% in income inequality between farmers and manufacturing workers.This surge in rural-urban inequality was not in line with official policy objectives, as laid out in Presi-dent Hu Jintao’s Harmonious Society program during the mid-2000s. In 2017, the 13th five-year planintroduced targeted subsidies to alleviate poverty among tobacco farmers.11 Such transfer schemesmay not have been necessary in the absence of a consolidation. Although the policy’s official objec-tive was to increase efficiency by exploiting scale economies, I do not find strong evidence for eitherincreasing returns to scale or other scale economies.

The principal contribution of this paper is to study the effects of ownership consolidation on bothmarkups, markdowns and productivity together. There is a large empirical literature on market powerand ownership consolidation, such as Miller and Weinberg (2017); Nevo (2001), which mostly buildson the ‘demand-side’ approach of S. Berry, Levinsohn, and Pakes (1995). In principle, this approach

10High markdowns are also consistent with widespread poverty among Chinese tobacco farmers, in contrast to most othertobacco-growing countries where tobacco ranks high among crops in terms of profitability (FAO, 2003).

11https://www.tobaccoasia.com/features/china-aims-to-increase-tobacco-farmers-income/

4

can be used to identify the effects of consolidation on both markups, markdowns and productiveefficiency. Doing so requires imposing a model on how firms compete on product markets and inputmarkets, and identifying both the product demand and input supply curves. Moreover, observableand exogenous variation in market structure is necessary to identify efficiency gains. In practice,applications using this approach have tended to assume either input prices or efficiency gains to beexogenous to ownership consolidation.

Alternatively, the production and cost approach to markups of Hall (1986); De Loecker and Warzyn-ski (2012) can be used as well. If all inputs are substitutable, then this approach can be used to identifyboth markups, markdowns and productive efficiency without imposing a model of competition on ei-ther input or product markets, but at the cost of having to identify the production function.12 I show,however, that joint identification of markups and markdowns using this approach alone fails wheneverfirms have buying power over non-substitutable inputs. In many other industries, a subset of interme-diate inputs are not substitutable, or substitutable to a limited extent. Examples include beer brewing(hop), coffee roasting (beans) and consumer electronics (rare earth metals).13

I solve this identification challenge by combining the production and cost approach with a modelof input supply. This allows identification of markups, markdowns and productive efficiency withoutimposing a model of how firms compete on cigarette markets. This modeling choice is driven by theinstitutional setting of the Chinese tobacco industry: it is far easier to model leaf supply than cigarettedemand. Tobacco leaf is sold on static spot markets that are isolated by law. In contrast, cigarettemarkets are not well-delineated, conduct is unknown due to the government-owned wholesaler, retailprices are unobserved, and cigarette demand is dynamic due to addiction. In other settings, however,it may be preferred to combine a model of product demand and production, or a model of productdemand and input supply.

I find that ignoring non-substitutability in the production and cost approach leads to mis-inferenceof both markups and markdowns. Markups are biased upwards when inputs are erroneously assumedto be substitutable. In the substitutable inputs production model, marginal costs can be expressed forinput separately, and depend on the input price and the markdown. In the Leontief production model,however, the prices and markdowns of all inputs enter marginal cost, because all inputs need to bemoved together to change output. Markdowns are, conversely, underestimated. The reason for this isthat an output elasticity below one will be imposed for the Leontief input. I illustrate these sources ofbias in my application: assuming substitutable tobacco leaf delivers markup and markdown averages

12This was shown in De Loecker, Goldberg, Khandelwal, and Pavcnik (2016) and applied in, among others, Morlacco(2017); Brooks, Kaboski, Yao, and Qian (2019)

13Even if intermediate inputs are partially substitutable, but to a lower degree than implied by a Cobb-Douglas productionfunction, the implications of this paper still matter. Markdowns and markups would then be weakly identified, ratherthan non-identified. One could, for instance, think about settings in which firms can substitute in-house production ofintermediate inputs with outsourcing.

5

of 5.63 and 0.77, while these are 1.06 and 2.87 when using the Leontief production function.

I make three additional contributions to the literature. First, I contribute to the literatures aboutefficiency gains from consolidation (Braguinsky, Ohyama, Okazaki, & Syverson, 2015; Blonigen &Pierce, 2016; Grieco, Pinkse, & Slade, 2017) and State Owned Enterprise (SOE) reform and privatiza-tion (Gupta, 2005; Hsieh & Song, 2015; Chen, Igami, Sawada, & Xiao, 2018). These papers usuallyfind evidence for large TFP gains from SOE privatization. Hsieh and Song (2015) finds that consol-idation policies similar to the one studied in this paper led to an increase in aggregate TFP of 20%across all Chinese manufacturing industries. My contribution to this literature is to allow for endoge-nous input prices, which changes the interpretation of the productivity residual. Using a productionfunction with exogenous leaf prices leads to the conclusion that average TFP increased by 30% due tothe consolidation. In reality, leaf prices fell as monopsony power increased on leaf markets. This hasimportant policy implications: large-scale consolidation policies, such as the one studied in this paper,are increasingly implemented. China consolidated many of its SOEs into industrial giants in variousimportant industries such as energy, transport utilities, telecommunication and defense industries.14

Other countries, such as Indonesia, have recently adopted similar policies as well. If these reformslead to rising monopsony power rather than productivity growth, this changes their welfare effects.

Second, I contribute to the literature on monopsony power, such as Card and Krueger (1994);Manning (2003).15 Recent papers on this topic includes Matsudaira (2014) and Wollmann (2019) fornurses, Naidu, Nyarko, and Wang (2016) for migrant workers in the UAE, and Goolsbee and Syverson(2019) for US academics. These papers usually estimate an input supply function. Valid input pricesinstruments are hence crucial, but hard to find.16 I propose using estimated productivity shocks fromthe production model as input demand shifters to identify the input supply curve, and discuss theconditions under which the underlying exclusion restriction holds.

Third, I contribute to the literature on rural-urban income inequality in developing countries. Manypapers have been devoted to this margin of inequality in China, as it has increased rapidly since theearly 1990s (Yang, 1999; Benjamin, Brandt, & Giles, 2005; Ravallion & Chen, 2009). My con-tribution consists of showing that SOE consolidation and the ensuing rise in monopsony power onagricultural markets was an important driver of this surge in inequality, at least in the tobacco indus-try.17 In a closely related paper, Chatterjee (2019) examines the consequences of interstate agriculturaltrade restrictions on farm income. In contrast to that paper, I focus on a different policy measure, SOE

14These policies are known as “Grasping the large and letting the small go” (Naughton, 2007).15A review with empirical work is in Ashenfelter, Farber, and Ransom (2010).16S. Berry, Azar, and Marinescu (2019) gives an overview of potential instruments that can be used.17While the relationship between market power and income inequality has been studied before, e.g. in De Loecker and

Eeckhout (2018); De Loecker et al. (2018), their focus has been mainly on product market power rather than on monopsonypower.

6

consolidation reforms.

The remainder of this paper is structured as follows. In the next section, I briefly summarize theindustry background and the data, and provide reduced-form evidence on how prices were affected bythe consolidation. In section 3, I present the model and discuss identification of markups, markdownsand productivity. Section 4 contains the estimation and the main results, while the distributionalconsequences and scale benefits are discussed in section 5. I revisit the main modeling assumptionsand present some extensions in section 6, after which I conclude.

2 The Chinese tobacco industry

2.1 Industry background

This paper focuses on the production of cigarettes in China, the value chain of which is visualizedin figure 1. Cigarette manufacturing firms, which are the entity observed in the data, buy tobaccoleaf from farmers at ‘purchasing stations’. In 1997, there were around 20 million tobacco farmers inChina, which were mostly organized at the household level and operated small plots of around 0.3-0.4 ha (FAO, 2003). Tobacco leaf needs to be ‘cured’ before it can be consumed, and various curingprocesses are possible.18 Farmers must sell their leaf at their nearest purchasing station. Only officialcigarette manufacturing firms are allowed to purchase tobacco leaf at these stations, and transportingtobacco leaf beyond these isolated markets without official approval is strictly forbidden by law (StateCouncil of the People’s Republic of China, 1997). Cigarette manufacturers, of which there were 150-350 during the time period studied, can buy at more than one station, but only within the boundaries oftheir own provinces. Tobacco leaves are sorted into quality ‘grades’, each of which sells at a differentprice.

[Figure 1 here]

Four out of five manufacturers in the dataset are owned by county, prefectural or provincial gov-ernments. All manufacturing firms are formally part of a state-owned holding, the Chinese National

Tobacco Corporation. In practice, they are autonomous in how they operate and set input prices andhow they compete against each other (Peng, 1996; Wang, 2013). SOEs are also controlled by local orprovincial governments, rather than the central government. They sell their cigarettes to wholesalerswhich are controlled by the State Tobacco Monopoly Administration (STMA) through its commer-

18Examples include air curing, fire curing and flue curing.

7

cial counterpart, the Chinese National Tobacco Trade Corporation (CNTTC).19 This organization iscentrally controlled and operates a monopoly on the cigarette market. In contrast to tobacco leaf,cigarettes are transported within and sold throughout China (State Council of the People’s Republicof China, 1997). The distinction between centrally controlled wholesaling and decentralized manu-facturing has been at the core of the STMA system since its inception in the early 1980s.

Even after China acceded to the WTO in 2001, the Chinese tobacco industry has been shieldedfrom international competition. Industry-wide exports and imports were merely 1.0% and 0.2% oftotal industry revenue between 1998 and 2007 (United Nations, 2019). The fiscal importance of thetobacco industry may be an important reason for this protection: in 1997, tobacco taxes and monopolyprofits made up for 10.4% of central government revenue. In 2015, tax revenues from the cigarettesindustry amounted to ¥840 B, which is 6.2% of China’s total tax revenue (Goodchild & Zheng, 2018;State Administration of Taxation, 2015).

Cigarette production process

Cigarette factories turn ‘cured’ tobacco leaf, paper and filters into cigarettes using labor and capital.Intermediate inputs make up for 90% of variable input expenditure, and tobacco leaf accounts fortwo thirds of intermediate input expenditure.20 Manufacturers treat, shred and compress the leaf, andthen insert it into cigarettes, cigars or other tobacco products. While the focus of the analysis willbe on cigarette manufacturers, I also include other tobacco users such as cigar and chewing tobaccoproducers in the market definitions, as these compete for leaf as well. They account for less than 5% ofindustry revenue, however. A picture of the consecutive steps in the cigarette production process arein panels (b)-(d) of figure 1. The focus will be on firms, rather than individual factories, as cigarettefactories are usually located in urban areas, often in the prefectural seat.

A map of tobacco manufacturing locations in China is in figure 2, with dots indicating countieswith at least one manufacturing firm. The tobacco industry is scattered across the country, even if themost important tobacco provinces, such as Yunnan, Guizhou, Sichuan and Henan, are in the south.Comparing the map of locations in 1999 and 2006 reveals that the number of counties with at leastone firm decreased drastically over this period of time. In 1999, there were 339 firms in 268 counties.By 2006, 160 firms were left in 121 counties.

[Figure 2 here]

19STMA and CNTTC share most of their leadership (Wang, 2013)20The Chinese data do not break down intermediate inputs into more detailed categories, but US census data from 1997

show that tobacco leaves make up for for 60% of all intermediate input costs in tobacco manufacturing firms (U.S. CensusBureau, 1997b)

8

Farmers

Tobacco farming is not a very profitable profitable activity in China. In 1997, just before the sampleperiod of this paper, tobacco was the median cash crop in terms of profitability (FAO, 2003). By2004, however, it had become the least profitable cash crop (Hu et al., 2006). Although tobaccofarmers can switch to other crops, this entails large switching costs. A policy intervention in whichtobacco farmers were helped to substitute crops in 2008 found that substituting increased annualrevenue per acre by 21% to 110% (Li et al., 2012). The fact that farmers do not substitute despitethese potential gains implies that crop switching costs are large. Farmers can also exit agriculturealtogether, but rural mobility is constrained due to the Hukou registration system. Some sources alsomake mention of tobacco farmers being coerced not to switch crops by local politicians, due to theimportance of tobacco for local fiscal revenue (Peng, 1996). Land tenure insecurity does, finally, alsomake migration more costly. Rural land is the property of villages or collectives, and if householdsmove they lose their exclusive use rights (Minale, 2018). This makes migration more costly and risky.

2.2 Data

For the main analysis in this paper, I combine two datasets. First, I use firm-level production and costdata between 1999 and 2006 from the Annual Survey of Industrial Firms, which is conducted by theNational Bureau for Statistics (NBS). I retain all firms with CIC codes 1610, 1620 and 1690, whichtogether correspond to HS code 24, “Tobacco and Manufactured Tobacco Substitutes”. This resultsin 478 firms and 2139 observations. The above-scale survey includes non-SOEs with sales exceeding5 million RMB and all SOEs irrespective of their size. I refer to Brandt, Van Biesebroeck, and Zhang(2012) for a comprehensive discussion of this dataset. Second, I obtain product-firm-month levelproduction quantities during the same time period, again from the NBS. Quantities are observed foronly 1,260 observations and 274 firms.21 In addition to these data, I match county-level populationcensus data and brand-level cigarette characteristics to the dataset as well. I use this data for therobustness checks. More details about all datasets used and data cleaning is in appendix A.

21Some sample selection may be going on due to missing quantities. Firms for which quantities are unobserved have onaverage less employees. The labor and material shares of revenue are, however, not significantly different between firmswith and without observed quantities. Whether quantities are observable explains barely any variation in revenue shares.

9

Summary statistics

Summary statistics of selected variables are in table 1. The average firm earned a revenue of $105million (in 2006 US dollars) and sold 340,000 cases per year. The average factory-gate price for a caseof 50,000 cigarettes was $1623, so the price for a pack of 20 cigarettes was on average $0.65. The 5thand 95th percentiles of pack prices were $0.02 and $1.16. Using retail price data from Nargis et al.(2019), this means that factory-gate prices were on average around 20% to 30% of retail prices, andthe difference between both includes wholesale margins, retail margins, transport costs and taxes.22

[Table 1 here]

2.3 Reduced-form evidence: consolidation and prices

The consolidation

The number of tobacco manufacturing firms fell from 351 in 1998 to 148 in 2007. In its annual reportfrom 2000, the STMA decided that “competitive large enterprise groups” had to be formed, withoutspecifying a concrete timing (Wang, 2013).23 The main motivation for this policy was to “enableChina’s cigarette industry to achieve scale and efficiency” (STMA, 2002). The top chart in figure 3shows that the number of manufacturers indeed started to decrease from that time onwards. In May2002, the STMA published a concrete implementation plan which ordered firms producing less than100,000 cigarette cases per year to be closed down in 2003,24 while firms with an annual productionbelow 300,000 cases were encouraged to merge with larger firms.25

The second graph in figure 3 compares the number of firms which produce less and more than100,000 cases per year. As quantities are observed for only a subset of firms, the annual number offirms reported is lower compared to the previous graph. The number of firms under the exit thresholdfell sharply between 2002 and 2004, from 98 to 25, while the number of firms above the thresholdsfell from 98 to 66. These smaller firms were economically meaningful: 36% and 46% of firms pro-duced less than 100,000 and 300,000 cases respectively in 2002, generating 6% and 11% of industry

22Harris (1998) reports that US wholesale prices were 59% of retail prices in 1998. Assuming a similar retail margin forChina would leave 29% to 39% of the retail price as profit for the wholesaler. This is a large margin, which is consistentwith the STMA’s monopoly power in wholesaling.

23I test for announcement effects in appendix C.124One case contains 50,000 sticks of cigarettes (Fang, Lee, & Sejpal, 2017)25The thresholds were calculated based on production in 2002. In appendix C.1, I test for bunching just above the exit

threshold, and find no evidence for this.

10

revenue. As figure 3 shows, average market shares of the provincial market leaders increased frombelow 50% to 70% between 2002 and 2007. For the two largest firms in each province, joint marketshares increased from 70% to 85%.

[Figure 3 here]

Of all firms that produced below the exit threshold, 22 firms continued to exist after 2003. Amongthese firms, 12 were not state-owned, so the STMA could not force them to close down. A further 7were dropped during the data cleaning procedure due to anomalies such as negative intermediate inputexpenditures. Merely 3 SOEs below the exit threshold survived after 2003, for unknown reasons.

Appendix table A2 shows the ‘first stage’ regressions of how input market Hirschman-Herfindahlindices changed differently between firms with and without competitors below the exit threshold in2002. In markets which were subject to the consolidation policy, the HHI indices for both materialsand labor increased by a third compared to control markets when using province-level market defini-tions. At the prefecture-level, the concentration indices increased by a fifth, and at the county level bya one seventh.26

Factor revenue shares

Panel (a) of figure 4 plots the evolution of total labor and intermediate input expenditure over totalrevenue in the industry (all deflated). The aggregate labor share of revenue fluctuated at around 3%,while the aggregate intermediate input share of revenue fell from 40% to 25% between 2000 and2007. Changes in relative input expenditure can be due to various reasons other than market power,such as technical change. I will come back to these reasons in the next section.

[Figure 4 here]

26The higher increase in concentration at the province-level masks the fact that in many counties, the number of firmsdropped to zero, which means that these markets are omitted from the analysis. Accounting for these markets, the consol-idation was actually more pronounced the more narrowly markets are defined.

11

Treatment and control groups

To what extent did consolidation contribute to the fall in the intermediate input share of revenue? Letthe number of firms i producing less than 100,000 cases Q per year t in market i be denoted Nit:

Nit =∑f∈i

(I[Qft < 100, 000])

Firms producing less than 100,000 cases were forced to exit in 2003. I construct a consolidationtreatment variable Cf which is a dummy indicating whether firm f is located in a county in whichthere was at least one firm producing below the exit threshold in 2002, at the onset of the size-dependent consolidation.

Cf = I[Ni,2002 > 0]

Panel (a) of table A1 contains some more information about the treatment and control group sizes.Before the policy was implemented in 2003, half of the firms produced less than 100,000 cases,and together earned 8.1% of total industry revenue. The treatment group consists of 15% of the firmsproducing above the threshold that had at least one competitor below the threshold in the same county.At the prefecture level, 39% of firms were in the treatment group.27 The revenue and output shares ofthe treatment group were similar to the share of the number of firms.

I compare cigarette prices, leaf prices and production quantities between the treatment and controlgroups before 2003 in panel (b) of table A1. When defining input markets at a more disaggregatedlevel, the pre-treatment differences in terms of prices between treatment and control groups are larger.Firms in treated counties already charged lower cigarette and leaf prices before the policy started thanfirms in the control group. At the province-level, this was not the case. This may have to do withannouncement effects, which I explore in appendix C.1.

Difference-in-differences model

In order to assess the effects of consolidation on input and product prices, I estimate the difference-in-differences model in equation (1). I compare firms with and without competitors below the exitthreshold before and after 2003. The outcome of interest yft is product prices pft, leaf prices wMft ,and wages wLft; all in logs. I also use the log factor revenue shares as the outcome variable. The labor

revenue share is defined as αLft ≡WLftLft

PftQftand the intermediate input revenue share as αMft ≡

WMftMft

PftQft,

with labor and intermediate input units L,M , cigarette prices P and quantities Q. The consolidationdummy Cf itself is not included as it is part of the firm fixed effect θyf . The coefficient of interest that

27When defining input markets at the province-level, 86% of firms are in the treatment group.

12

quantifies the consolidation effects is θy2 .

yft = θy0 + θy1I[t ≥ 2003] + θy2CfI[t ≥ 2003] + θy3t+ θyf + υyft (1)

with yft ∈ pft, wMft , wLft,WMftMft

PftQft

,WLftLft

PftQft

Pre-trends

The usual assumptions for a difference-in-differences model apply. First, the error term υft has to beconditionally independent from the consolidation dummy CfI[t ≥ 2003]. The error term υy containsall time series variation in outcome yft that is not captured by the other control variables. Second, thetrends in the outcome variable need to be parallel for both treatment and control groups in the absenceof the treatment.

The main question underlying this exclusion restriction is why there were firms operating belowthe threshold in some markets, but not in others. Figure A1 shows that most firms were distributedclose to the exit threshold of 100,000 cases before 2003. Panel (b) of table A1 shows that firms in thetreatment and control groups were comparable in terms of leaf and cigarette prices and size before2003.

I verify the parallel trends assumption by testing whether the treatment and control groups hadparallel trends in the outcome variables of interest before 2003. I estimate equation (2) on the timeperiod 1999-2002. The coefficient of interest is the interaction effect between the treatment variableand time, η2: if this coefficient is close to zero and insignificant, parallel pre-trends in the outcomesof interest cannot be rejected.

yft = ηy1Cf + ηy2Cf ∗ t+ ηy3t+ νyftif t < 2003 (2)

The estimates of equation (2) are in panel (d) of table 2 and confirm that it cannot be rejected that thepre-trends for both wages, leaf prices and cigarette prices were parallel before 2003.

Results

I start with visual evidence. Panel (b) of figure 4 compares the evolution of average relative inputexpenditure between the treatment and control group. The average ratio of intermediate input expen-diture over the wage bill fell from 11 to 8 for firms in the treatment group between 2002 and 2006.For firms in the control group, it increased from 11 to 12. Taking the weighted averages by laborusage or the median, in panels (c)-(d), yields very similar patterns.

13

The drop of relative intermediate input expenditure can be due to two reasons. One explanationcould be factor-biased technological change. For tobacco leaf, this is unlikely: the required amount oftobacco leaf per cigarette varies little across firms.28 A factor-biased shock that changed the amountof labor needed per cigarette is possible. In order to generate the patterns in figure 4, however, the re-quired amount of labor per cigarette would have to increase over the sample period. Mechanization ofproduction has the opposite effect. Secondly, these patterns could be due to a change in relative inputprices. This could be either due to increased monopsony power on tobacco leaf markets, decreasedmonopsony power on labor markets, or both.

I examine this further by estimating equation (1) using cigarette prices, leaf prices and wages peremployee as the outcome variable. I estimate the same model using different input market definitionsat the province, prefecture and county level in panels (a), (b) and (c) of table 2.29 I find that in allthree specifications, labor wages fell by 14.7%, 12.0% and 5.4% respectively,30 but this drop wasnot statistically significant. Leaf prices did, in contrast, fall by 37% when defining leaf markets atthe province or prefecture level, and by 53% when using county-level leaf markets. Cigarette pricesdid, finally, decrease by 23-32%, depending on the leaf market definition used. The narrower mar-kets are defined, the larger the effects. Consolidation hence led both to lower leaf prices and lowerfactory-gate cigarette prices. The relative drop in leaf prices was significantly larger than the drop incigarette prices at the prefecture and county level. The effect of a firm exiting in the same county islarger compared to a firm exiting at the other side of the same province, as manufacturers are likelyto compete more for leaf with nearby manufacturers.

[Table 2 here]

Variation in prices due to consolidation, which is arguably exogenous in this setting, is not sufficientto draw conclusions about the underlying mechanism. Falling leaf and factory-gate prices could bedue to increased markdowns and increased markups, but changes in productive efficiency would alsolead to different equilibrium input and product prices.31 I therefore combine the natural experimentwith a model of production and input market competition.

28More evidence on this is in appendix C.29In panel (d), I test the parallel pre-trend assumption using equation (2). The time trends in all three outcome variables

were not significantly different before the policy was implemented, so parallel pre-trends cannot be rejected.30exp(−0.160)− 1 = −0.14731Changes in productivity alone could not explain the combination of falling input prices and falling product prices together,

though.

14

3 A model of production, markups and markdowns

3.1 Production and costs

Production

Cigarette manufacturers f produce Qft cases of cigarettes using a quantity of tobacco leaf Mft, laborLft and fixed assetsKft.32 I allow for substitution between labor and capital, but not between tobaccoleaf and labor or capital. Tobacco leaf may be substitutable to a limited extent with capital, forinstance due to waste reducing technologies. The crucial substitution pattern when inferring markupsand markdowns is, however, the substitutability between leaf and labor.33 I assume it is impossible .34

Let the production function be given by equation (3):

Qft = minβMftMft,ΩftH(Lft, Kft)

exp(εft) (3)

The amount of tobacco leaf needed to produce a case of cigarettes is given by βMft . In the baselinemodel, I impose that firms do not differ in terms of leaf content, βMft = βM .35 Firms do differ interms of their productivity level Ωft. In line with most of the prior literature, this productivity termis assumed to be a Hicks-neutral scalar. Firms use a common production technology H(.) whichparametrizes the substitution pattern between labor and capital. I assume H(.) is twice differentiablein both labor and capital. Measurement error in output is denoted εft. Equation (3) nests productionfunctions in which all inputs are substitutable: the input requirement βM would then be zero bydefinition, and intermediate inputs added as a substitutable input.

Even if tobacco leaf is not substitutable in the cigarette production function, total intermediate in-puts could be substitutable with total labor if it were possible to vertically integrate cigarette factorieswith farms. This is not the case in the Chinese tobacco industry (Peng, 1996; FAO, 2003; Wang,2013). If firms have large buying power over intermediate input suppliers, they also have less incen-tives to integrate vertically.

32Other intermediate inputs, such as paper and filters, are also part of M . I abstract from these in the model as they togetherrepresent less than a third of intermediate input costs and are as non-substitutable as leaves.

33The reason for this is that the first order cost minimization conditions are derived for leaf and labor, not for the capitalstock which is assumed fixed.

34Dunn and Heien (1985) found no substitution between agricultural inputs in food processing industries. Sumner andAlston (1987) argued tobacco leaf is substitutable with non-farm inputs, which goes against my Leontief model. Theyused, however, an input demand approach which rules out monopsony power. If leaf prices are endogenous, regressingrelative input usage on relative input prices will naturally result in a positive correlation, even if inputs are not substitutable.I will estimate the substitution elasticity between all inputs in section 6.

35I will extend the model to allow for both factor-augmenting productivity and heterogeneous leaf content in section 6.

15

The capital stock evolves dynamically with a depreciation rate δK and investment Ift:

Kft = δKKft−1 + Ift−1

I assume firms produce a single but differentiated product, cigarettes, at price Pft.36 Cigarettemarkets are allowed to be imperfectly competitive, so the price of cigarettes is a function of howmuch a firm and its competitors produce. I deal with quality differences in cigarettes and tobacco leafby using a price control in the production function, as in De Loecker et al. (2016).37

Input markets

I assume that both tobacco leaf and labor are variable and static inputs, meaning that they can be ad-justed without adjustment costs and are proportional to the number of cigarettes produced38 Tobaccoleaf is sold at least at monthly intervals at the purchasing stations, and more cigarettes require moretobacco. While the NBS survey does not separately report production and non-production workers,the US census does. In 1997, over 70% of US cigarette manufacturing employees were productionworkers, and hence variable (U.S. Census Bureau, 1997a).39 I choose to model labor as a static vari-able as hiring and firing flexibility is considerably higher in China compared to the global and regionalmedian countries in terms of labor market flexibility (World Economic Forum, 2019).

The prices for labor and tobacco leaf are denoted WLft and WM

ft . They depend on exogenous firmcharacteristics, some of which are observed, denoted Zft, and some of which are latent, denoted ζft.Examples of these characteristics are whether a firm is state-owned or not (observed) and how closeit is located to a highway (latent). Next, input prices potentially also depend on how many inputs thefirm uses in a given year. If this is the case, then the supply curve for input V ∈ M,L is upward-sloping. Finally, equilibrium input prices also depend on characteristics and input prices of all otherfirms in the same market i. This set of firms is denoted as Fit.

W Vft = W V (Zft, Vft, ζft,Z−ft, V−ft, ζ−ft) ∀ − f ∈ Fit

The price elasticity of input supply is defined as ψVft − 1 and can differ across firms and over time. IfψV equals one, the input supply function is flat, which means that the price of input V is exogenous

36The model can be generalized to a multi-product setting by using De Loecker et al. (2016), but this is not of first-orderimportance for the tobacco industry as the average firm earns more than 90% from selling cigarettes.

37More details and the exact empirical specifications used follow in section 4.38I follow the classification of Ackerberg, Caves, and Frazer (2015). In section 6.2, I discuss inventories and dynamic leaf

demand.39These workers accounted for 65% of the wage bill.

16

to the firm.

ψVft ≡∂W V

ft

∂Vft

VftW Vft

+ 1 ≥ 1 for Vft ∈ Mft, Lft (4)

I allow for input market power on the market for tobacco leaf, so ψM can be above one. I assume,however, that firms are price takers on labor and capital markets: ψL = 1. Manufacturing workersare much more mobile than farmers in China, and can also easily switch to other manufacturingindustries, as their skills are not specific to cigarette manufacturing. Tobacco farmers, in contrast,can only supply cigarette manufacturers or switch crops, which is very costly. I will provide moreempirical evidence for this assumption in section 6. The method could easily be generalized to allowfor endogenous wages WL

ft .40

As I estimate the input supply model using data with one-year intervals, a short-run supply elasticityis inferred. I discuss the difference between short- and long-run supply elasticities in section 6.2.

Decision-making

I assume manufacturing firms choose the tobacco leaf price WMft each year in order to minimize

per-period variable costs, taking the other input prices as given. In theory, the Chinese governmentcentrally regulated leaf prices and procurement up to 2015, after which leaf markets were liberalized(Lan, 2015). In reality, however, manufacturing firms had considerable pricing power on leaf marketsfrom the 1980s onwards. Peng (1996) provides anecdotal evidence for this: leaf prices were, forinstance, set for 15 ‘quality grades’.41 As these grades were mostly subjective, manufacturing firmscould set prices by choosing the corresponding grade. Peng (1996) mentions frequent conflicts be-tween peasants and manufacturers over grading, bitter feelings among the farmers and, in some cases,peasants being forced to sell tobacco at a price below their cost of production.

The assumption that firms minimize costs can be questioned. It is often suggested that state-ownedenterprises (SOEs) have non-profit objectives such as generating local employment (Lu & Yu, 2015).In the tobacco industry, Peng (1996) notes that cigarette manufacturers have “the purpose of makingprofits” and “often bargain with each other for better deals”.42

Assumption 1. — Firms choose input pricesWMft annually to minimize a short-term cost function,

taking the other input prices as given.

40This would require to impose a similar structure on how firms compete on labor markets to the model of intermediateinput competition of this paper.

41I cope with this input quality differentiation in the production model by adding a price control, similar to De Loecker etal. (2016).

42In appendix B.5, I still extend the model to allow for objective functions other than cost minimization. Different objectivefunctions will change the inferred markup levels. As the vast majority of tobacco manufacturers are SOEs anyway, it isunlikely that the observed changes in markups and markdowns will be driven by differences in firm objectives.

17

The associated Lagrangian of the cost minimization problem is given by equation (5), with marginalcosts λft:

WMft

∗= arg minLft = WM

ftM(WMft )+WL

ftLft+λft

(Qft−Qt(M(WM

ft ), Lft, Kft,Ωft, βMft ))

(5)

As labor and tobacco leaf cannot be substituted, there is just one first order condition, rather thanone for each variable input.43 When firms choose the leaf price, and hence the quantity of tobaccoleaf used, they automatically also choose how much labor to use.

3.2 Markups

General case: endogenous input prices and non-substitutable inputs

The markup is defined as the ratio of cigarette prices Pft over marginal costs λft:

µft ≡Pftλft

Solving the first order condition in equation (5) for marginal costs yields the markup expression inequation (6a), which is derived in appendix B.1.

µft =(αLftβLft

ψLft + αMftψMft

)−1

(6a)

with αVft ≡VftW

Vft

PftQftεft

for V ∈ L,M

The output elasticity of labor βLft is retrieved when estimating the production function. The revenueshares αVft on the right-hand side of equation (6a) are in principle observed, but I follow De Loeckerand Warzynski (2012) and adjust for measurement error ε. This means that I calculate revenue sharesusing quantities in which the estimated error ε is netted out. As said before, labor wages are exogenousin this paper: (ψLft − 1) = 0. Even then, the right-hand side of equation (6a) still contains the inputprice elasticity of tobacco leaf, ψM , which is latent. The intuition for the fact that markups depend onthe input supply elasticity is that the slope of the input supply curve is part of marginal costs: if thefirm increases output by one unit, costs increase by more the steeper the input supply function is, asinput prices endogenously increase.

Markups µft and the input price elasticity (ψMft − 1) are hence not separately identified without

43This also applied to the beer brewing production function in De Loecker and Scott (2016).

18

adding more structure to the model. Even if the firm would have more variable inputs with exogenousprices, this does not lead to separate identification of ψ and µ. As none of these inputs would besubstitutable with tobacco leaf, the input demand conditions for all these inputs would incorporate theendogenous price effect in the same way.

Special case (i): Exogenous input prices and substitutable inputs

Suppose all inputs have exogenous prices and are mutually substitutable. In that case, the non-substitutable input requirement is by definition zero, βMft = 0, and all input price elasticities areall zero: (ψV − 1) = 0 , ∀V . The markup expression then simplifies to the formula from De Loeckerand Warzynski (2012):

µft =βLftαLft

(6b)

Special case (ii): Endogenous input prices and substitutable inputs

Next, consider a setting in which all inputs are substitutable, but in which input prices are endogenous.This implies upward-sloping input supply functions. The markup is now expressed as the outputelasticity of a variable input divided by its revenue share and divided by the input price elasticity ofsupply plus one. This corresponds to the expression from Morlacco (2017).

µft =βLft

αLftψLft

(6c)

If there is just one variable substitutable input L, markups and markdowns are not separately iden-tified. In case there are multiple variable substitutable inputs of which at least one input has anexogenous input price, markups and markdowns can be separately identified. Suppose, for instance,that there are two variable inputs L and Lc, and that the price of Lc is given from the firm’s perspec-tive. Equation (6c) then becomes a system of 2 equations. Because the markdown LC is assumed tobe one, there are two unknowns: the markup µ and the markdown under the price of input L.

µft =βLft

αLftψLft

µft =βL

c

ft

αLc

ft

The markdown can now be found by dividing both equations. This is the markdown expression fromMorlacco (2017):

ψLft =αL

c

ft βLft

αLftβLcft

19

Special case (iii): Perfect input markets and a non-substitutable input

A final special case holds when all input prices are exogenous, but when there is one input that cannotbe substituted for any other input. In this case, βMft > 0, but all input supply elasticities (ψVft − 1) arezero. The markup is given by equation (6c), which corresponds to De Loecker and Scott (2016). It isidentified even if there is only one substitutable input.

µft =(αLftβLft

+ αMft

)−1

(6d)

Note that cases (ii) and (iii) can be blended: if the substitutable input prices are endogenous, butnon-substitutable input prices are not, the markup is identified as long as there is at least one substi-tutable input with an exogenous price.

Possible approaches to achieve identification

Back to the general case in equation (6a). In order to separately identify the markup and the inter-mediate input supply elasticity, additional structure needs to be imposed on how firms compete eitheron product or input markets. One could, for instance, identify markups µft by imposing a modelof how firms compete on product markets and back out the input supply elasticity without furtherassumptions. In this paper, I take the opposite approach and identify the input price elasticity of leafsupply by imposing a model of how firms compete on leaf markets. The reasons why I prefer thissecond approach are entirely specific to the context of Chinese cigarette manufacturing. Leaf marketsare easier to define as they are isolated by law, while cigarette markets are not. Moreover, it is harderto model cigarette demand than leaf supply. Demand for cigarettes is likely to be dynamic due toaddiction. Moreover, the wholesaler’s conduct and retail prices are unobserved. Even if there is aliterature that addresses each of these complications,44 modeling leaf supply is easier as leaf is soldon well-delineated static spot markets.

44Dynamic demand is modeled in general by Gowrisankaran and Rysman (2012) and specifically for cigarettes consumptionby Baltagi and Levin (1986). There is also a literature on conduct identification (Ciliberto & Williams, 2014; Miller &Weinberg, 2017; Michel & Weiergraeber, 2018). Delipalla and O’Donnell (2001) estimate conduct of European cigarettemanufacturers.

20

3.3 Markdowns

Markdown interpretation of the input supply elasticity

The input price elasticity of supply, (ψMft − 1), is intimately related to the concept of the ‘markdown’,the wedge between an input’s marginal revenue product and its price. To see this, the cost minimiza-tion problem in equation (5) must be rewritten as a profit maximization problem. Using the derivationin appendix B.2, the following expression is obtained:

ψMftWMft =

∂(PftQft)

∂Mft︸ ︷︷ ︸MRP leaf

− βMftWLft

∂Lft∂Qft︸ ︷︷ ︸

Marginal labor cost

(7)

The right-hand side of this equation is the marginal revenue product of tobacco leaves (‘MRP leaf’)minus the marginal labor cost of increasing the number of tobacco leaves by an additional unit. Thisentire term can hence be thought of as the marginal profit gain from an additional unit of leaf, netof leaf costs. The term ψM can now be interpreted as a ‘markdown’: it is the ratio of the marginalbenefit of leaves divided by the leaf price farmers receive. If the markdown is large, farmers get asmall fraction of their contribution to manufacturing profits.

Input competition model

Let there be It isolated leaf markets i. Farmers j can sell tobacco leaves to manufacturing firmsf ∈ Fit, with f = 0 indicating the outside option of not selling to any firm. I assume each firmoperates in exactly one market and that farmers sell their entire production to a single firm, whichmakes sense as there were 20 million household-level farms producing tobacco leaf in 1997 (FAO,2003), selling to merely 350 firms. I follow S. T. Berry (1994) and the ensuing literature and let theutility function of an farmer j depend on the leaf price, manufacturer characteristics Zft and ζft andan i.i.d. type-I distributed manufacturer-farmer utility shock νjft. I allow for the fact that farmers maynot accept the highest bid, as it is possible that non-price characteristics ζ influence farm utility aswell. The location of the manufacturing plant or who owns the factory could, for instance, influencefarmer utility conditional on the leaf price offered.

Ujft = γWWMft + γZZft + ζft + νjft

I normalize the utility of the outside option to zero, as usually: Uj0t = 0. I explicitly rule out variationin supplier preferences for input prices or manufacturer characteristics, γW and γZ . In appendix C.4,

21

I allow for more flexible substitution patterns by using a nested logit model.45 In this section, I alsoestimate alternative input model specifications, such as a logs-on-logs model.

I assume that the farmer-manufacturer utility shock νjft is i.i.d. across firms, farmers and time, andfollow the usual logit assumption:

Assumption 2. — The farmer-manufacturer utility shock νjft follows an extreme-value type-Idistribution.

As stated before, firms choose tobacco leaf prices simultaneously in order to minimize per-period

costs. I denote the tobacco leaf market share of firm f in year t as Sft =Mfjt∑

r∈FitMfrt

. The distribu-

tional assumption on νjft allows the intermediate input share Sft to be written as:

Sft =exp(γWWM

ft + γZZft + ζft)∑r∈Fit exp(γWWM

rt + γZZrt + ζrt)

Dividing this share by the market share of the outside option S0t, whose utility is normalized tozero, as well as taking logs, leads to equation 8, which will be estimated in the next section.

sft − s0t = γWWMft + γZZft + ζft (8)

The markdown (ψ) can, finally, be expressed as a function of observable input prices and input marketshares, and of the estimated valuation parameter γW :

ψMft ≡( ∂Sft∂WM

ft

WMft

Sft

)−1

+ 1 =(γWWM

ft (1− Sft))−1

+ 1 (9)

4 Empirical analysis

The empirical analysis consists of four main steps. I start by estimating the production function. Sec-ondly, I estimate the input supply function in order to retrieve the markdown. Third, I combine theoutput elasticity estimates from the first step with the markdown estimates from the second step tocalculate markups. Fourth, I examine how the 2003 consolidation policy affected both markdowns,markups and productivity. Finally, I revisit some of the key modeling assumptions in some exten-sions. The estimated output elasticities are used both when estimating the input supply function andwhen estimating markups. I therefore bootstrap the entire estimation procedure using 50 bootstrapiterations.

45When applying the methods used to labor markets, heterogeneous supplier preferences and all kinds of other frictionsseem especially important.

22

4.1 Production function identification

Identification of the production function serves three purposes. First, the output elasticity of labor isrequired to identify the markup. Next, it allows quantifying returns to scale. Finally, the productivityresiduals will be used as leaf demand shifters in order to identify the input price elasticity of leafsupply.

If input prices would be exogenous and the production function a constant returns to scale Cobb-Douglas production function, the cost shares of each input would simply equal their output elasticity(Syverson, 2004).46 In the context of this paper, the Cobb-Douglas function in labor and capitalonly contains inputs with exogenous input prices. I do not want to impose constant returns to scale,however, as a key motivation of this paper is to allow for increasing returns to scale and efficiencygains from consolidation.47

I therefore estimate the production function. Tobacco leaf can be ignored due to the Leontiefassumption: only the H(.) function needs to be estimated.48 Denoting logs of variables in lowercases,equation (10a) needs to be estimated.

qft = h(lft, kft) + ωft (10a)

A dummy Cft ∈ 0, 1 indicates whether a firm is subject to the consolidation policy, which will bediscussed later in more detail. I allow productivity to endogenously change due to consolidation, asin De Loecker (2013); Braguinsky et al. (2015):

ωt = g(ωft−1, Cft) + ξft

Product and input differentiation

The log production function in equation (10a) has both input and output quantities on both sides ofthe equation. Cigarettes are, however, differentiated products with various quality levels. As pointedout in De Loecker et al. (2016), this can result in bias as firms which produce high-quality, high-pricecigarettes use more labor per unit of cigarette than low-quality producers. Moreover, I observe thenumber of employees l rather than the total amount of hours worked l, and the skills of workers,

46If, however, monopsony power differs across firms, this no longer holds.47In appendix C.3, I impose constant returns to scale and use the cost shares approach to find the output elasticities of labor

and capital, and use these to estimate markups and markdowns.48In general, it could be optimal for firms to diverge from the Leontief ‘first order condition’ of intermediate inputs equaling

the H(.) function in labor and capital, as argued in Ghandi, Navarro, and Rivers (2018) The assumption that intermediateinputs enter the production function linearly solves this problem, however.

23

which could differ across firms. The unit wage WL, which is the wage bill divided by the numberof employees, captures variation in both hours worked and worker skills. I follow De Loecker et al.(2016) by adding a function a(.) in both product cigarette prices and unit wages to the productionfunction. The production function in logs, equation (10a), is hence replaced by equation (10b):

qft = h(lft, kft) + a(wLft, pft) + ωft (10b)

Latent markups and markdowns

The usual approach to identify the production function in the presence of a latent productivity scalarhas been to invert an input demand function and express the productivity residual ω as a function ofobservables only (Olley & Pakes, 1996; Levinsohn & Petrin, 2003; Ackerberg et al., 2015). Follow-ing the production function, equation (3), leaf demand depends on productivity, the leaf content percigarette and the optimal amount of labor and capital used:

Mft =Ωft

βMH(Lft, Kft)

If intermediate input quantities are observed, identification of the production function would bestraightforward. As shown in appendix A of Ackerberg et al. (2015), the first stage inversion canbe achieved by regressing quantities on labor, intermediate inputs and capital. The only reason whyintermediate input usage differs across firms, conditional on labor and capital used, is variation inproductivity Ωft.

Intermediate input quantities are, however, latent. It is hence necessary to invert the labor demandequation for productivity. When solving the first order condition in equation (5), firms choose a leafprice that minimizes variable costs. An optimal output level Q∗ft corresponds to this leaf price. Firmschoose the mix between labor and capital that minimizes costs

As derived in appendix B.3, labor demand depends on cigarette prices, other inputs and their prices,ownership consolidation, and firm characteristics Z such as its export status or ownership structure,which are all observed. It also depends, however, on markups µ, markdowns ψM and productivityωft, which are all latent. If two of these three unobservables are latent, which is very likely, the inputdemand inversion breaks down. If firms use few inputs, we cannot be certain whether this is due tolow productivity, high markups or high markdowns.

lft = l(kft,WMft ,W

Lft,Zft, ωft, µft, ψ

Mft )

24

Solution 1: Input prices and market shares in input demand

The first strategy to identify the production function is to stick to the inversion approach, but to insertdrivers of markdown and markup variation across firms in the input demand function. This approachwas used in De Loecker et al. (2016). In most models of supply and demand, markup and markdownvariation across firms depends on prices and market shares. In the logit model outlined in section3.3, for instance, markdown variation depends on the tobacco leaf price and leaf market share. Ifthe same holds on the cigarette market, then including leaf and cigarette prices and upstream anddownstream market shares in the input demand function should solve the problem. Defining marketshares downstream can be difficult and was one of the reasons that the demand estimation approachwas not followed, but we can insert multiple market shares at different geographical levels (county,prefecture, province) in the demand function.

In appendix D, I discuss the validity of this approach in more detail using simulated data. I find thatthe production function can be consistently identified using this approach if prices are observed andmarket definitions correctly defined, under the condition that there is no unobserved heterogeneityin supplier and consumer valuations of prices and firm characteristics. As soon as farmers differ,for instance, in their valuation of firm characteristics, such as whether it is state-owned or not, theproduction function estimates can be seriously biased. The intuition for this is that in this case,markdowns differ across firms even after conditioning on input market shares and input prices, asfirms take into account these heterogeneous valuations when choosing inputs. Because of this, latentmarkdowns enter as a persistent unobservable into the input demand function, which is no longerinvertible in latent productivity.

I follow the two-stage procedure of Ackerberg et al. (2015). In the first stage, I invert the leafdemand function for the latent productivity term. This results in equation (11). The vector of demandshifters zft includes all observables which may affect input demand: product and ownership types,export dummies and export shares of revenue in the vector of input demand shifters zft. I use a third-order polynomial in labor and capital in Φ(.), and include the leaf price and input market share toproxy for markdown variation, as discussed in the previous section.

qft = Φ(lft, kft, wLft, w

Mft , zft, pft, sft) + εft (11)

In the baseline model, I use a Cobb-Douglas function in labor and capital for H(.). In appendixC.2, I estimate a Translog production function instead, which results in very similar estimates.

h(Lft, Kft) = βLlft + βKkft

In line with the timing assumptions imposed for capital and labor, the moment conditions to identify

25

the output elasticities are given by:

Eξft(β

L, βK)

(lft−1

kft

)= 0

Solution 2: BLP instruments

An alternative identification strategy is to exploit input price variation across firms. This was alreadyexplored in an older productivity literature, as in Griliches and Mairesse (1998). The productionfunction literature has stepped away from using input prices as instruments because it is doubtfulthat they are orthogonal to total factor productivity.49 I rely on an instrumental variables approachinstead. 50 When input markets are oligopsonic, however, a whole new set of possible instruments forinput prices and quantities become available, similar to the oligopoly demand identification literature(S. Berry et al., 1995).

In a recent paper, S. Berry et al. (2019) discuss the use of instruments to identify input supplyin oligopsony. If these instruments can be used to instrument input prices when identifying the inputsupply curve, then they can also be used to instrument input quantities when identifying the productionfunction. Characteristics of competing cigarette manufacturers in the same leaf market could, forinstance, be used as shifters of the leaf price, and hence leaf quantities, provided that these firmcharacteristics enter farmer utility or shift input demand.

Consider manufacturer characteristics which enter the farmer utility function (Zft, Zft) ∈ Zft,such as the firm’s ownership structure. The characteristics of competing firms in the same market isdenoted Xft:

Xft =1

|Fit| − 1

( ∑r∈Fit

(Zrt)− Zft

)If there are as many suitable characteristics as coefficients for the substitutable inputs (in this paper,

2), the moment conditions are:

Eωjt(βl, βk)

(zjt

zjt

)= 0

In order for the exclusion restriction to be satisfied, the selected characteristics z of competing firmsneed to be orthogonal to the firm’s productivity level, conditional on all observables. There are mul-tiple threats to this identification strategy. First, if there are productivity spillovers across firms or

49Moreover, there is not much input price variation within narrow industry-market pairs in countries with centralized wagebargaining.

50A more detailed discussion on the use of instruments in production function identification is in Ackerberg, Benkard, Berry,and Pakes (2007).

26

X-inefficiencies, this would violate the exclusion restriction. Secondly, firm characteristics that arestrategically chosen as part of the input market game played by the oligopsonists are not to be used asinstruments.

I select two sets of variables as the instrument Z. First, I use import tariffs in other manufacturingindustries in the same county, sales-weighted across firms in the county, as an instrument. Even iffirms cannot set labor wages themselves, equilibrium wages will be affected by input demand in othermanufacturing industries. The disadvantage of this instrument is that it is county- and year-specific,not firm-specific. Three out of four cigarette manufacturers are, however, the only producer in theircounty. Secondly, I add the average export share of revenue of competitors and export participationof competitors in the same prefecture as instruments. As competitors enter the international cigarettesmarket, their cigarettes demand increases, and hence also leaf prices and optimal input quantities. Ifthere are no ‘productivity spillovers’ from exporting to non-exporting firms, export shocks at otherfirms would be excluded from the own firm’s productivity level. As exports make up less than 1% oftotal revenue, this is unlikely to be a strong instrument.

In appendix D, I detail further the identification of production functions when input prices areendogenous, and use simulated data to compare the performance of both the IV and control functionapproaches.

Estimates

The estimated output elasticities are in panel (a) of table 3. The left columns contains the estimateswhich are inferred when using ACF: the output elasticities of labor and capital are 0.320 and 0.761respectively. The standard errors are large, as is usual when estimating ACF on a small sample.51 Thescale parameter is 1.052, which implies modest scale returns. The scale parameter is not significantlydifferent from one, but this may be a result of a lack of power due to the small sample.

The right columns report the estimates when using export participation and export shares of com-peting firms and the weighted tariff rate in other manufacturing industries in the same leaf market asinstruments for labor and capital. The firm’s export status and export share are controlled for in theproduction function. The estimated output elasticities of labor and capital are now 0.404 and 0.699.The scale parameter of 1.104 suggests modestly increasing returns to scale, but has a relatively largestandard error. The sample using IV estimation is larger than when using ACF, as lagged variablesare not needed.52 Throughout the rest of the paper, I will use the IV estimates.53

51This was, for instance, also the case in the industry-by-industry estimates from De Loecker et al. (2016)52In panel (b) of table A11, I re-estimate the IV model on the ACF sample size. Both output elasticities using the IV model

are now much closer to the ACF estimates, at 0.365 and 0.804.53I conduct different robustness checks by using the ACF estimates, a translog production function and a substitutable leaf

27

[Table 3 here]

4.2 Markdown and markup estimation

Input supply estimation

Next, I estimate the input supply function, equation (8). I control for three observable firm character-istics Zft which are likely to affect leaf supply. First, I control for cigarette prices, as they are a proxyfor high quality. Secondly, I control for manufacturer ownership dummies. As the Chinese tobaccoindustry is highly regulated and politically influenced, the utility from selling to a manufacturer maydiffer depending on whether this manufacturer is state-owned, collectively owned or a private firm.Finally, I include prefectural dummies in order to control for the manufacturer’s location.

It is unlikely that controlling for these observable supply shifters suffices to identify the leaf supplyfunction. If manufacturers have unobservable characteristics ζft which enter farmer utility, then thiswill affect the equilibrium leaf price.54 This results in the same kind of simultaneity bias as whenestimating product demand. In order to separately identify input demand and supply, an instrumentfor input prices is hence needed. This has to be a shifter of input demand which does not enter inputsupply, i.e. does not enter farmer utility.

The productivity residual ωft, which was estimated in the previous section, is such an input demandshifter. In order for the exclusion restriction to hold, productivity of the cigarette manufacturer shouldbe excluded from farmer utility, conditional on the leaf price. In other words, conditional on howmuch they are paid, farmers do not care how efficient the manufacturing firms they are selling to are.

There are various ways in which this exclusion restriction could be violated. First, it could bethat manufacturers prefer to buy repeatedly from the same suppliers, for instance due to switchingcosts or asymmetric information. In this case, farmers could prefer to sell to highly productive firmswhich are less likely to exit than less productive firms. Secondly, selling to a highly productive firmcould be a signal for high quality produce if there is asymmetric information about leaf quality. Byimposing the exclusion restriction, I rule out these mechanisms. The first mechanism cannot be ruledout a priori, but anecdotal evidence from Peng (1996); Wang (2013) depict leaf markets as monthlyauctions at purchasing points, at which it does not seem to be costly for manufacturers to change

production function in appendix C.54Examples of such characteristics are whether the factory is easily accessible by road, how long it has been producing,

political connections of its managers, etc.

28

suppliers. Moreover, leaf quality is examined prior to the purchase being made, as prices are set infunction of the quality ‘grade’. Asymmetric quality information therefore does not seem of first orderimportance in this industry.55

Second, consolidation dummies Cf , which were defined earlier in section 2.3 are candidate in-struments. Operating in a market subject to the consolidation affected the input market shares andbuying power of manufacturers, and hence their equilibrium leaf demand. If the consolidation did notenter farmer utility, conditionally on the leaf price, then these consolidation dummies can be used asinstruments for leaf supply.

I use these consolidation dummies to test for overidentifying restrictions in appendix table A5.The Sargan-Hansen χ2 is 0.988. If the consolidation dummies are excluded instruments, it cannot berejected that manufacturer TFP is a valid instrument for leaf prices as well.

Once the input supply elasticity is estimated, markdowns ψM are inferred by using equation (9). Fi-nally, I use both the output elasticity of labor and the markdown to calculate markups, using equation(6a).

Measurement

As is usually the case in production and cost datasets, I do not observe leaf prices WMft , but rather leaf

expenditure MftWMft . Because of the Leontief production function, dividing leaf expenditure by the

output quantity results in the leaf price divided by the leaf concentration βMft .

MftWMft

Qft

=WMft

βMft

In the baseline model, I assume all cigarette producers have the same tobacco content per cigarette,βM . Additional data reveal very little variation in leaf contents per cigarette across firms. Moreover, aslong as firms did not adjust the tobacco content in response to consolidation, the estimated treatmenteffects should not change.

Market definitions

Leaf markets need to be defined in order to estimate the leaf supply function. There is usually atleast one purchasing station per county. As became clear from the map in figure 2, however, many

55When applying the same model on manufacturing labor markets, more caution is needed. There are many reasons whyemployees would prefer to work for highly productive firms, even if these offer lower wages, such as career dynamics orbetter working conditions.

29

counties ended up without a tobacco manufacturing firm, even though tobacco farming continued.The market definition hence has to be wider than the county-level at least for a subset of firms. Themarket definition also has to be less than or equal to the province-level, as leaf trade across provinceswithin China is restricted. In the baseline approach, I define input markets at the province-level,because this yields the most conservative estimates for the treatment effects. Firms never compete onleaf markets outside their own province due to the transport restrictions. I do, however, re-estimatethe main regressions with prefecture- and county-level leaf market definitions as robustness checks.Input supply becomes more inelastic the more local input markets are defined, implying higher leafprice markdowns.

Input supply estimates

The estimates of the leaf supply function, equation (8) are in panel (b) of table 3. The OLS esti-mates are in the first column, and indicate a downward-sloping supply curve. As already mentioned,however, this is likely due to simultaneity bias. The IV estimates are in the second column of panel(b), and imply an upward-sloping supply curve. The average leaf price per pack is RMB 1.12. Thecoefficient implies that an increase in this leaf price per pack of RMB 0.10, a relative increase of 9%,leads to an increase in the leaf market share of 44%. Productivity residuals are a strong instrument, asthe first stage F-statistic is 79.14. Converting these estimates to input supply elasticities using equa-tion (9) delivers an average inverse supply elasticity of 0.54.56 A survey of estimated tobacco leafsupply elasticities in Askari and Cummings (1977) reports similar elasticities: 0.71 for India, 0.51 forBangladesh and 0.60 for Nigeria.

Throughout the paper, I have assumed that manufacturing labor wages are exogenous. Althoughthis assumption was motivated in the industry background section, it can be tested empirically. Iestimate the elasticity of the labor supply function in the same way I did for tobacco leaf. I usethe same specification and instruments as for tobacco leaf, with the exception that I add exports ofcompetitors as a second instrument, because the first stage is weaker than for tobacco leaf prices.The results are in panel (c) of table 3. The left columns report the OLS estimates and the rightcolumns report the IV estimates. The estimated elasticity of labor supply is very close to zero andstatistically insignificant both with and without instruments. This supports the assumption made thatmanufacturing labor wages are exogenous in this industry.

56The average markdown is 2.869. The average input supply elasticity is hence 1/(2.869− 1) = 0.535

30

Markdown and markup estimates

I report selected moments for both markdowns and markups in panel (a) of table 4. The first twocolumns correspond to the baseline model in which tobacco leaf and other inputs cannot be sub-stituted. The average markdown is 2.87, which means that farmers receive 35% of their marginalcontribution to firm profits, according to (equation 7). Markups are on average 1.06, which impliesthat cigarette prices surpass marginal costs by 6.2%. This combination of large manufacturing mar-gins on leaf markets and small margins on the wholesale market is consistent with the the fact thatthey buy from many small farmers, but sell to a monopsonic wholesaler. The standard errors on themarkdown and markup estimates are due to uncertainty about both the estimated input supply elastic-ity and output elasticities of inputs. The average markdown lies significantly above one, which is thelevel that implies exogenous input prices, while the average markup is not significantly different fromone.

[Table 4 here]

The entire distribution of markups and markdowns is plotted in figure 5. Around half of manu-facturing firms sell to the CNTTC at prices below their marginal costs, which does not mean theyare loss-making, as they still have positive markdowns. The markdown distribution lies above one,meaning that all manufacturing firms have at least some input market power on the leaf market. Thereis a long tail of firms with markdowns above 3, meaning that farmers receive less than one third oftheir marginal revenue product.

[Figure 5 here]

Comparison to the substitutable input model

How important was it to model leaf as being non-substitutable? In panel (b) of table 4, I compare themarkdown and markup estimates using the more conventional model with substitutable tobacco leaf,which is also used in Morlacco (2017). Markups are now estimated to be on average 5.63, but with avery large standard error. Markdowns are, to the contrary, estimated to be 0.77, implying that farmersare paid more than their marginal profit contribution. Given the industry background, these markupand markdown estimates are unrealistic.

The markup expression when inputs are substitutable, µSUBft , is an overestimate of real markups if

31

there is a non-substitutable input. This can be seen by comparing markup expressions (6b) and (6a):

µSUBft =βLftαLft

>(αLftβLft

+ αMftψMft

)−1

if αMft > 0

The intuition for this overestimation is that in the non-substitutable model, marginal costs incorporateboth the cost of intermediate inputs and labor: both have to be increased simultaneously in order toincrease output (De Loecker & Scott, 2016).

The markdown in the substitutable inputs model, ψM,SUBft is obtained by dividing the markup over

the input with exogenous prices, labor, over the markup using the input with endogenous prices,tobacco leaf, as explained in Morlacco (2017):

ψM,SUBft =

βMftβLft

αLftαMft

This expression is wrong in any case in the non-substitutable inputs model, as the first order conditionsfor tobacco leaf and labor are not linearly independent. Still, the substitutable production function willgenerally underestimate the markdown as the estimated output elasticity of leaf βM will be belowunity, while it is one in reality due to non-substitutability of tobacco leaf. The substitutable inputsmodel compares the revenue share of materials to its output elasticity, and will hence underestimatethe difference between both.

Comparison to prior monopsony studies

How large are the markdown estimates implied by this paper? Most prior studies of monopsonypower estimate the input price elasticity. I convert these estimates into the ratio of input prices overmarginal revenue products using equation (9), which is on average 0.35 in this paper. Appendix tableA14 compares this ratio to a selection of prior studies about monopsony power. A lower input priceto marginal revenue share implies more monopsony power. The ratios range from 0.83 for tenure-track professors in US academia (Goolsbee & Syverson, 2019)57 to 0.70 for female grocery clerks inSouthern USA (Oaxaca & Ransom, 2010).

The estimated degree of monopsony power on Chinese tobacco leaf markets is hence much largercompared to these prior studies. One reason relates to the outside option of the input suppliers.Tobacco farmers sell to a narrow set of cigarette firms, which makes their earnings sensitive to changesin downstream market structure. As discussed in the industry background, tobacco farmers face largecrop switching costs, for instance due to crop-specific capital and skill requirements. Supply of inputs

57Monopsony power was estimated to be higher for associate and full professors.

32

that are less industry-specific is more likely to be elastic. Skill requirements for grocery employeesseem, for instance, not specific to grocery stores. Depending on their fields of research, it is likelythat college professors can be employed outside of academia.

A second key difference lies in the degree of labor mobility costs. Artuc, Lederman, and Porto(2015) estimate labor mobility costs to be higher in China compared to the USA and Germany. WithinChina, mobility costs are higher in rural areas because of the Hukou permit system and land-rightsuncertainty. The more costly it is for farmers to migrate in order to switch industries, the more monop-sony power this gives to manufacturing firms buying their products.

More research is needed to confirm the importance of monopsony power on labor markets in otherindustries and other countries. The approach used in this paper offers a framework to conduct similarresearch in these other settings. The bulk of research on monopsony power still focuses on the USA.In many countries throughout the developing world, labor mobility costs are higher than in China(Artuc et al., 2015), which makes the existence and importance of monopsony power more likely.

Markdown and markup drivers

Which firm and county characteristics explain variation in markups and markdowns? In panel (c) oftable 4, I report some correlations. Both markups and markdowns are increasing in relation to firmsize, although the relationship is stronger for markdowns. State-owned enterprises have higher mark-downs, but not higher markups. Counties with higher unemployment rates have higher markdownlevels: worse outside options for farmers imply higher monopsony power for cigarette producers.Counties with high migration rates, defined as the share of the population that was born in anothercounty, have much lower markdowns. Higher migration rates imply a more mobile labor force. Firmscannot decrease leaf prices too much in these counties, as workers would migrate more rapidly in re-sponse. Finally, corporate tax rates, defined as the share of revenue which is taxed by the government,are higher for high markdown firms, but not for high markup firms. This suggests that governmentsextract part of the rents generated by monopsony power, besides taking their share of firm profits inSOEs.

4.3 The effects of consolidation on markups, markdowns and productivity

With the markup, markdown and productivity estimates at hand, I now examine how these wereaffected by the ownership consolidation. I use the difference-in-differences model shown earlier,equation (1). Rather than prices, I now use markups, markdowns and productivity as the outcome

33

variables y.

The main identifying assumptions were already outlined in section 2.3. First, the time series varia-tion in the outcomes, conditional on all control variables, should be independent from whether firmshad competitors below the exit threshold in 2002 or not. Second, the pre-trends in all outcome vari-ables should be parallel for both treatment and control groups.

Consolidation treatment effects

The treatment effect estimates from equation (1) are in panel (a) of table 5. I use province-level marketdefinitions.58 Markdowns increased by 37% for firms affected by consolidation compared to firms inthe control group.59 Markups did, in contrast, fall by 23%. Productivity is estimated to have increasedby 8% more for firms in consolidated markets, but this difference is not statistically significant.

The rise in markdowns and fall in markups is consistent with the fact that manufacturers face manysmall suppliers, but a single monopolistic buyer. The manufacturer surplus which was generatedby market power gains on the leaf market seem to have been partly extracted by the government-controlled wholesaler, which exerted a downward pressure on factory-gate cigarette prices. In thenext section, I will examine in more detail who benefited and lost the most from the consolidationpolicy.

[Table 5 here]

Treatment effects in the substitutable leaf model

Earlier in this section, I showed that markup and markdown levels were not realistic when estimat-ing a production function with substitutable tobacco leaves. In panel (b) of table 5, I compare theconsolidation treatment estimates when using a Cobb-Douglas model in labor, leaves and capital tothe Leontief model of the previous paragraph. The markup and markdown effects are estimated tobe 25% and -14% compared to the Leontief model, but have the same sign. Both models interpret adeclining revenue share of intermediate inputs and increasing revenue share of labor as evidence forincreased markdowns and decreased markups.

58In appendix table A21, I re-estimate the treatment effects in equation (1) when defining input markets at the prefectureand county level. As markets are defined more narrowly, treatment effects on markups and markdowns increase, whilethe TFP gains remain of similar magnitudes. Firms located in the same county are more likely to be direct competitors fortobacco leaf compared to firms located at the opposite side of the same province.

59= exp(0.314)− 1

34

The TFP treatment effects are, in contrast, estimated to be 35% when using the substitutable inputsmodel, compared to 8% in the Leontief model. The reason for this difference is that input pricesand input quantities are not separately observed. De Loecker and Goldberg (2014) discussed howunobserved input quantities led to biased production function coefficients when inputs differ in termsof quality. The source of bias in this paper is, in contrast, monopsony power. A drop in intermedi-ate input prices will be interpreted as rising productivity when not accounting for endogenous inputprices:

qft = βLlft + βKkft + βM(mft + wMft ) + ωft

In the Leontief model, intermediate inputs do not enter the estimated production function, andhence unobserved leaf prices do not enter the productivity residual. Prior work on SOE privatizationand consolidation policies found that they led to large increases in profitability (Gupta, 2005; Brown,Earle, & Telegdy, 2006; Hsieh & Song, 2015; Chen et al., 2018). These profitability gains could bedue to both increased monopsony power or TFP growth.

Pre-trends

I again test whether markups, markdowns and productivity had parallel trends between the treatmentand control groups before the treatment started by estimating equation (2). The estimated interactioneffects of the time trends and treatment categories from equation (2) are in panel (c) of table 5. Paral-lel pre-trends before 2003 cannot be rejected.

Heterogeneous effects

I interact the consolidation treatment effects on markdowns in equation (1) with firm and county char-acteristics, in order to see where markdowns increased the most, controlling for province dummies. Inpanel (a) of table 6, I find that the markdown increase was larger in counties in which there were morefirms under the exit threshold of 100,000 cigarette cases in 2002. The change in market structure waslarger where there were more firms producing under the exit threshold. I also find that the markdownincrease was largest for the smallest firms. As the exiting firms were mainly small, this exit was arelatively larger shock for smaller firms.

In panel (b), I interact the treatment effects with demographic characteristics of each county. Inthe left columns, I find that the markdown increase was larger in counties with higher unemploymentrates in 2000. Farmers had less attractive outside options in high unemployment areas, which allowedmanufacturers to increase the leaf price markdown. In column 2, I interact the treatment effect vari-able with the migrant share of the county population in 2000. Markdowns increased more in counties

35

with relatively more immigrants. Both unemployment and migration are, however, partially decisionsmade by workers. These correlations therefore do not necessarily imply causality.

[Table 6 here]

5 Distributional consequences and scale economies

5.1 Rural-urban income inequality

What were the distributional consequences of the consolidation policy? I focus on income inequalitybetween manufacturing workers and tobacco farmers, which are based mainly in urban and rural areasrespectively. Rural-urban income inequality has risen sharply in China over the past two decades(Yang, 1999; Benjamin et al., 2005; Ravallion & Chen, 2009). This has also been the case in thetobacco industry: while the average manufacturing wage grew by 14.5% per year between 1999 and2006, average tobacco leaf prices fell by 5% per year on average. Accounting profits increased byno less than 24% per year for cigarette manufacturers. These patterns show that two margins ofinequality increased: first, the income gap between manufacturing workers and farmers rose sharply,and second, the gap between firm profits and employee wages rose as well.

Tobacco leaf prices are not the only determinant of income for farmers, as production quantitiesand farm productivity are as well. Aggregate producer statistics from the Food and Agriculture Orga-nization (FAO) show, however, that farm sizes remained constant and yields per acre grew by merely1.8% per year during this time period (FAO, 2019), not enough to compensate falling leaf prices.

Back-of-the-envelope calculation

Equation (7) quantified the relationship between the leaf price and the markdown. The marginalgain from an additional unit of leaf equals the marginal revenue product of leaf MRPM minus themarginal labor cost that is needed for this, MLC. The leaf price equals the ratio of this gain over themarkdown:

WMft =

MRPMft −MLCftψMft

I now conduct a back-of-the-envelope calculation of the extent to which the consolidation policycould have contributed to income inequality between tobacco farmers and manufacturing workers. Ilet markdowns of firms in the treatment group evolve in the same way as they did for firms in the

36

control group. I assume both the marginal revenue product of leaf and marginal labor costs to haveremained unchanged. Using notation from equation (1), the alternative leaf price WM becomes:

WMft =

WMft exp(θψ

M

2 ) if t ≥ 2003 & Cf = 1

WMft otherwise

(12)

Figure 6 compares the evolution of leaf prices to manufacturing wages between 1999 and 2006,with both series being normalized at 0 in 2003. Before 2003, manufacturing wages (blue line) alreadyoutgrew leaf prices (red solid line). Between 2003 and 2006, manufacturing wages increased by 60%,while leaf prices fell by 10%. The dashed red line shows that without enforcing the exit thresholds,leaf prices would have grown by 20% over this time period. The consolidation hence explains 40%of the increase in income inequality between cigarette manufacturing workers and tobacco farmers.

[Figure 6 here]

Caveats

The analysis above is by no means a counterfactual simulation. First, it ignores entry and exit. Higherentry and lower exit of farmers could have led to different equilibrium leaf prices. Second, the analysisis of a partial equilibrium nature. As tobacco represents a large share of economic activity in someprovinces, changes to leaf prices would also have affected equilibrium cigarette prices, manufacturingwages and prices and wages in other sectors.

5.2 Margins along the value chain

I now compare the effects of the consolidation on the surplus of all actors throughout the value chain.In equation (13), I decompose the cigarette retail price P ret, which is the total value generated inthe industry, into the share that farmers, manufacturers, and both wholesaler and retailers receive.60

Farmer income is assumed to be entirely generated from tobacco leaf, and farm input prices arenormalized to zero. The variable profit margin of manufacturing firms and their employees receiveis the difference between the factory-gate cigarette price and the leaf price. The wholesaler/retailerreceives, finally, the difference between the retail price and the factory-gate cigarette price.

60As I do not observe retail prices, I cannot distinguish wholesale from retail surplus

37

WM

P ret︸︷︷︸Farmer’s share

+P −WM

P ret︸ ︷︷ ︸Manufacturer’s share

+P ret − PP ret︸ ︷︷ ︸

Wholesaler & retailer’s share

= 1 (13)

As I do not observe firm-level retail prices, I use aggregate cigarette retail price data from Zheng,Wang, Hua, and Marquez (2016) to calculate average wholesaler profit margins. Retail prices could,however, have dropped in consolidated markets, which I cannot pick up using the aggregate data.

Retail price decomposition

In panel (a) of table 7, I decompose which share of the cigarette price is allocated to farmers, man-ufacturers and wholesalers/retailers,61 using equation (13). The leaf price is on average 29% of theretail cigarette price. The difference between the factory-gate price and the leaf price is another 35%of the retail price. The remaining 36% of value gets generated at the wholesaling and retailing level.

How were gross profit margins of these different agents affected by consolidation? In panel (b), Iestimate how these margins, calculated using equation (13), changed due to the consolidation treat-ment.62 Farmer gross profit margins fell by 39%, due to the drop in leaf prices, while manufacturingprofit margins increased by 10%. The profit gain from falling leaf prices is partially undone by thedrop in factory-gate cigarette prices. The wholesaler/retailer was, finally, the main winner from con-solidation: wholesaling gross profit margins increased by 31%. The consolidation policy was orderedand carried through by the wholesaler’s administrative counterpart, the STMA. It therefore seemslikely that consolidation served as a tool to increase profits of the wholesaling monopoly, rather thanas a means to increase manufacturing efficiency.

As firm-level retail prices are unobserved, I cannot distinguish wholesaler, retailer and consumersurplus. As the government-regulated cigarettes monopoly remained unchanged before and after2003, it seems likely that the wholesaler internalized the profits from lower factory-gate prices. Itcannot be ruled out, however, that this price drop was passed through to retailers and/or consumers.63

[Table 7 here]

61I do not observe wholesale prices, so cannot distinguish margins of wholesalers from retailers62I still use province-level market definitions.63Quantifying this pass-through would be crucial when examining the public health consequences of the consolidation

policy. If retail cigarette prices fell, cigarette consumption could have increased in response.

38

5.3 Returns to scale vs. scale economies

The official objective for enforcing the exit thresholds was to generate scale economies. In the discus-sion so far, I have mainly focused on returns to scale in production. Estimating the production functionrevealed that there were modestly increasing returns to scale, and that physical productivity slightlyincreased due to the consolidation, although these increases were statistically not very significant.

Forcing small manufacturers to exit may, however, have resulted in scale economies other thanincreasing returns to scale. Duplicated fixed costs may have been eliminated, for instance. I use theobserved profit levels Πft in order to examine these potential scale economies. Profits are definedas follows, with fixed costs FCft being all costs other than intermediate input or labor expenditure.Although I call these costs ‘fixed’, some variable costs, such as transport costs, may be part of thisterm as well.

Πft ≡ PftQft −MftWMft − LftWL

ft − FCft

Consolidation and scale economies

I examine whether total fixed costs per market evolved differently between markets with and withoutexit threshold enforcement. I re-estimate the difference-in-differences model, equation (1) at the levelof input markets i. I use the log total fixed costs per market i in year t, FCit, as the dependent variable.I use the indicator for the market having at least one firm below the exit treshold in 2002, Ci as theindependent variable and interact it with the post-treatment period:

log(FCit) = θF0 + θF1 I[t ≥ 2003] + θF2 CiI[t ≥ 2003] + θF3 Ci + υFit

I estimate this model at the province, prefecture and county level in table 8. In column 1, I use logfixed costs as inferred from the profit figures as the dependent variable. In column 2, I use the logcapital stock as the dependent variable. The sign of the treatment effects differs depending on themarket definition, and the treatment effects are never significant. There is hence no strong evidencefor scale economies generated by the consolidation policy.

6 Alternative explanations

I now revisit three key modeling assumptions. First, I test whether tobacco leaf and labor are reallynon-substitutable when producing cigarettes. Second, I revisit the assumption that the productivityshifter was Hicks-neutral in labor and capital. Third, I use data on brand-level product characteristicsto revisit the assumption that the tobacco content per cigarette was constant across firms.

39

6.1 Substitutable tobacco leaf

The non-substitutability between labor and tobacco leaf has been a maintained assumption throughoutthis paper. It can be tested empirically. Suppose the cigarette production function is no longer givenby the Leontief production function from (3), but by the following CES production function:

Qft =

((βMM

σM−1

σM

ft + βLLσM−1

σM

ft

) σM

σM−1

)βML

KβK

ft Ωft

The substitution elasticity σM measures the extent to which labor and tobacco can be substituted. Istill assume substitutability between variable inputs and capital.

Solving the first order conditions from equation (5) results in equation (14). Firms use relativelymore labor compared to tobacco leaf if wages are lower, if the output elasticity of labor is relativelyhigher, or if firms have more monopsony power over tobacco leaf.

lft −mft = σM(

ln(WMft )− ln(WL

ft))− σM

(ln(βM)− ln(βL)

)+ σM ln(1 + ψMft ) (14)

Estimating equation (14) is subject to the same simultaneity bias as when estimating the input sup-ply function. The extent of monopsony power of a firm affects its optimal input demand. Moreover,intermediate input prices are not observed separately. I use the same BLP instruments for wages whichwere introduced in the production function estimation section to estimate the elasticity of substitutionbetween labor and materials, σM .

The results are in column 1 of table A17. The elasticity of substitution between intermediate inputsand labor is estimated to be 0.21, but is not significantly different from zero, in line with the Leontiefmodel used throughout the paper. I can also re-estimate equation (14) with capital instead of tobaccoleaf on the left-hand side, which allows for flexible substitution between labor and capital, but notbetween labor and intermediate inputs. This also requires capital to be viewed as a variable input,which I do by means of a first approximation. The elasticity of substitution estimate between laborand capital is in column 2, and is estimated to be 0.89 with a standard error of 0.22. The unitarysubstitution between labor and capital which was imposed in the baseline Cobb-Douglas model hencecannot be rejected.

40

6.2 Static vs. dynamic demand and supply

Dynamic leaf demand

I model both leaf demand and supply decisions as being static. In reality, both are likely to have adynamic component as well. Manufacturers hold leaf inventories, which I abstract from in the currentmodel. If they dynamically optimize their input sourcing depending on the leaf price, this will changethe implied markup estimates.

Inventories can, however, only affect short-run markup fluctuations. I show in appendix table A10that the negative markup effect of the consolidation becomes smaller when limiting the time frame ofthe treatment effect from three to one year.

Dynamic leaf supply

The supply elasticities I estimated were short-run elasticities, as variation in prices and quantitiesusing one-year intervals was used to estimate the input supply function. It is likely that leaf supplyby farmers becomes more elastic when increasing the time horizon at which input supply is estimated(Hamilton, 1994). There are six months between sowing and harvesting, and it takes another two toeight weeks to cure the leaves.64 In the short run, sowing and curing costs are sunk, and farmers willnot exit as long as their variable profits are positive.

If tobacco leaf supply would be inelastic only in the short run, but elastic in the longer run, thenmarkdowns in consolidated markups should rise immediately after the exit threshold enforcement, butfall again later, as farmers start exiting. The evidence in appendix table A10 does, however, not alignwith this. The drop in markdowns in consolidated markets increased during the three years followingthe reform, rather than falling again. I conclude from this that leaf supply is not only inelastic in theshort run, but also over a longer period of at least three years.

6.3 Labor-augmenting productivity

The productivity shifter ω was assumed to be Hicks-neutral throughout the paper. What if there wasfactor-augmenting technical change? Equation (14) shows that factor-augmenting productivity andmarkdowns are not separately identified when using the production approach. Both factor-augmentingproductivity and monopsony power are inferred by comparing relative input usage in their respectiveliteratures. This potentially has implications outside of this paper. In a paper about migration in Chinausing the same dataset as this paper, Imbert, Seror, Zhang, and Zylberberg (2018) find that increased

64Source: https://www.pmi.com/glossary-section/glossary/tobacco-curing

41

immigration led firms to produce at a higher labor intensity. I find, however, that monopsony powerwas lower in counties with high migration rates, as workers are more mobile there. In a model whichtakes input prices as given, falling monopsony power would be picked up as lower capital intensity.

In the context of this paper, this is very unlikely to drive the stylized facts concerning input rev-enue shares. As the industry-wide relative cost share of labor increased relatively to leaf, this wouldmean that cigarette production became much less capital-intensive over time, in sharp contrast to thegeneral trend in Chinese manufacturing. Both the levels and changes in the estimated markups could,however, change when allowing for factor-biased technical change. I redefine the H(.) function inproduction function (3) to allow for labor-specific productivity ΩL and allow for flexible substitutionbetween labor and capital, measured by parameter σK . This is similar to the production function usedin Doraszelski and Jaumandreu (2017).

H(Lft, Kft) =(βKK

σK−1

σK

ft + βL(LftΩLf )

σK−1

σK

) σK

σK−1exp(ωft)

Using the same derivation as for equation (14) and treating capital as a variable input, labor-augmenting productivity can be estimated as the residual of the following equation, in which WK

is the interest rate. In contrast to equation (14) in the previous section, markdowns do not enter rela-tive input demand for capital and labor as the markdown level affects both input decisions to the sameextent. The relative input demand equation becomes:

kft − lft = σK(

ln(WLft)− ln(WK

ft ))

+ σK(

ln(βK)− ln(βL))

+ (1− σK)(ωLft)

The estimates of this equation were already shown in the second column of table A17. The markupformula, (6a) can be rewritten as follows, using the same derivation as before:

µft =( αLftβLftΩ

Lft

ψLft + αMftψMft

)−1

I re-estimate the difference-in-differences model, equation (1), using labor-augmenting productiv-ity as the left-hand side variable. The results are in appendix table A18. Labor-augmenting produc-tivity increased by 26%, 15% and 4% due to the consolidation when using province-, prefecture andcounty-level market, respectively.

In appendix table A13, I re-estimate the consolidation treatment effects on markdowns, markupsand total factor productivity, but add the log capital to labor ratio as a control variable. Controllingfor this variation in capital intensity across firms and over time does not change the treatment effectestimates by much.

42

6.4 Heterogeneous product characteristics

Finally, I revisit the assumption that the tobacco leaf content per cigarette, βMft , was constant acrossfirms. I use the product characteristics data on a subset of firms to show that in reality, there was lim-ited variation in tobacco concentration and in other cigarette characteristics, such as ventilation ratesand paper quality. The entire distribution of leaf contents per cigarette lies between 63% and 75%, sothe decline in the leaf share of revenue cannot be due to changes in the leaf content alone. Moreover,as long as product characteristics were similar between the control and treatment groups, they do notaffect the difference-in-difference estimates. Table A19 shows this is indeed the case: firms in thetreatment and control groups did not differ significantly in terms of any product characteristic, exceptthat they were less likely to be ‘ventilated’ cigarettes.

The second table in table A19 shows, however, that markdowns did not correlate with any productcharacteristic. Hence, the observed markdown increase for the treatment group cannot be due merelyto changes in cigarette design. Markups do correlate with product characteristics: cigarettes withhigher filter densities and more tobacco content are sold at higher markups, as these are more likelyto be of higher quality. The physical cost to produce cigarettes with different product characteristicsis very similar as both TFP and the leaf price are not significantly different across cigarette producttypes.

6.5 Further robustness checks

In appendix C, I carry out multiple other robustness checks. I narrow down the treatment time windowfrom 2004-2006 to 2004 only. I also compare exporting behavior across treatment and control groups,use different functional forms for the production function and include non-wage benefits as a variablelabor cost. None of these robustness checks alter the main findings to a meaningful extent.

7 Conclusion

In this paper, I examine the effects of ownership consolidation on buying power, market power andproductive efficiency. I find that a large-scale ownership consolidation in the Chinese tobacco in-dustry allowed cigarette manufacturers to increase the tobacco leaf price markdown by 37%. Thismonopsony power was both due to large crop switching costs, rural migration impediments, and alarge increase in downstream market concentration. For urban manufacturing workers, many of thesefrictions played a smaller role, and I find no evidence for monopsony power in manufacturing labormarkets. In response to the rise in monopsony power on leaf markets, monopsonic wholesalers re-

43

acted by increasing their own margins in the wholesaling market. Consolidation hence resulted notonly in lower leaf prices, but also in lower factory-gate cigarette prices. I find that markups of cigarettemanufacturers fell by 24%.

The main motivation to consolidate the tobacco industry was to increase scale economies. I findweak evidence for increasing returns to scale and for modest gains in total factor productivity. Takingendogenous input prices into account is crucial to interpret these productivity gains, and wronglyimposing exogenous input prices lead to the erroneous conclusion that productive efficiency increasedby 30%. I find no evidence for other types of scale economies, such as reduced fixed costs.

The consolidation policy had important distributional consequences. The main beneficiary of con-solidation was the government-controlled wholesaler, which monopolizes cigarette markets. Farmerincome was hit by the surge in monopsony power, and I find that forcing small manufacturers to exitexplained 42% of the rise in income inequality between tobacco farmers and manufacturing employ-ees between 2003 and 2006. As these groups mainly live in rural and urban areas respectively, theconsolidation policy contributed to the rise in urban-rural income inequality in the tobacco industry.

On the methodological level, I discuss the role of input substitution patterns for identification ofmarkups, markdowns and productive efficiency using the production and cost approach. I show thatlow input substitutability limits the identification of markups and markdowns using the productionand cost approach. I propose a solution to this challenge by complementing the production and costmodel with an input supply model.

Large-scale consolidation waves are a frequently used policy tool in transitioning countries, andthere is increased M&A activity in developed countries, such as the U.S. In order to fully understandthe effects of consolidation, it is crucial to take into account their effects on both productive efficiency,scale economies, and competition on input and product markets.

44

References

Ackerberg, D., Benkard, L., Berry, S., & Pakes, A. (2007). Econometric tools for analyzing marketoutcomes. In Handbook of econometrics (Vol. 6, p. 4171-4276).

Ackerberg, D., Caves, K., & Frazer, G. (2015). Identification properties of recent production functionestimators. Econometrica, 83(6), 2411-2451.

Artuc, E., Lederman, D., & Porto, G. (2015). A mapping of labor mobility costs in the developingworld. Journal of International Economics, 95(1), 28–41.

Ashenfelter, O. C., Farber, H., & Ransom, M. R. (2010). Labor market monopsony. Journal of Labor

Economics, 28(2), 203-210.Askari, H., & Cummings, J. T. (1977). Estimating agricultural supply response with the Nerlove

model: a survey. International economic review, 257–292.Autor, D., Dorn, D., Katz, L. F., Patterson, C., & Van Reenen, J. (2017). Concentrating on the fall of

the labor share. American Economic Review: Papers and Proceedings, 107(5), 180–85.Baltagi, B. H., & Levin, D. (1986). Estimating dynamic demand for cigarettes using panel data: the

effects of bootlegging, taxation and advertising reconsidered. The Review of Economics and

Statistics, 148–155.Benjamin, D., Brandt, L., & Giles, J. (2005). The evolution of income inequality in rural china.

Economic Development and Cultural Change, 53(4), 769-824.Berry, S., Azar, J., & Marinescu, I. (2019). Estimating labor market power.

Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica,64(4), 841-890.

Berry, S. T. (1994). Estimating discrete-choice models of product differentiation. The RAND Journal

of Economics, 242–262.Blonigen, B., & Pierce, J. (2016). Evidence for the effects of mergers on market power and efficiency

(Working paper No. 22750). National Bureau of Economic Research.Braguinsky, S., Ohyama, A., Okazaki, T., & Syverson, C. (2015). Acquisitions, productivity and

profitability: Evidence from the Japanese cotton spinning industry. The American Economic

Review, 105(7), 2086-2119.Brandt, L., Van Biesebroeck, J., Wang, L., & Zhang, Y. (2017). WTO accession and performance of

Chinese manufacturing firms. American Economic Review, 107(9), 2784-2820.Brandt, L., Van Biesebroeck, J., & Zhang, Y. (2012). Creative accounting or creative destruction?

firm-level productivity growth in Chinese manufacturing. Journal of Development Economics,97(2), 339-351.

Brooks, W. J., Kaboski, J. P., Yao, A. L., & Qian, W. (2019). Exploitation of labor? classical

monopsony power and labor’s share (Working Paper No. 25660). NBER.

45

Brown, J. D., Earle, J. S., & Telegdy, A. (2006). The productivity effects of privatization: Longitudinalestimates from Hungary, Romania, Russia, and Ukraine. Journal of Political Economy, 114(1),61–99.

Card, D., & Krueger, A. B. (1994). Minimum Wages and Employment: A Case Study of the FastFood Industry in New Jersey and Pennsylvania. American Economic Review, 84(4), 772-793.

Chatterjee, S. (2019). Market power and spatial competition in rural india (Working Paper No.2019/04). Cambridge INET.

Chen, Y., Igami, M., Sawada, M., & Xiao, M. (2018). Privatization and productivity in china.Available at SSRN 2695933.

Ciliberto, F., & Williams, J. W. (2014). Does multimarket contact facilitate tacit collusion? inferenceon conduct parameters in the airline industry. The RAND Journal of Economics, 45(4), 764–791.

Covarrubias, M., Gutierrez, G., & Philippon, T. (2019). From Good to Bad Concentration? US

Industries over the past 30 years (Tech. Rep.). National Bureau of Economic Research.Delipalla, S., & O’Donnell, O. (2001). Estimating tax incidence, market power and market conduct:

The European cigarette industry. International Journal of Industrial Organization, 19(6), 885–908.

De Loecker, J. (2013). Detecting learning by exporting. American Economic Journal: Microeco-

nomics, 5(3), 1-21.De Loecker, J., & Eeckhout, J. (2018). Global market power (Working Paper No. 24768). NBER.De Loecker, J., Eeckhout, J., & Unger, G. (2018). The rise of market power and the macroeconomic

implications (Working Paper). https://sites. google.com/site/deloeckerjan/research/RMPDLEU.pdf..

De Loecker, J., Goldberg, P., Khandelwal, A., & Pavcnik, N. (2016). Prices, markups and tradereform. Econometrica, 84(2), 445-510.

De Loecker, J., & Goldberg, P. K. (2014). Firm performance in a global market. Annu. Rev. Econ.,6(1), 201–227.

De Loecker, J., & Scott, P. T. (2016). Estimating market power: evidence from the us brewing

industry (Working Paper No. 22957). National Bureau of Economic Research.De Loecker, J., & Warzynski, F. (2012). Markups and firm-level export status. The American

Economic Review, 102(6), 2437-2471.Doraszelski, U., & Jaumandreu, J. (2017). Measuring the bias of technological change. Journal of

Political Economy, 126(3), 1027-1084.Dunn, J. E., & Heien, D. (1985). The demand for farm output. Western Journal of Agricultural

Economics, 10(1836-2016-150642), 13–22.Fang, J., Lee, K., & Sejpal, N. (2017). The China National Tobacco Corporation: From domestic to

global dragon? Global Public Health, 12(3), 315-334.

46

FAO. (2003). Issues in the global tobacco economy: Selected case studies. In Fao commodity studies

(Vol. 1810-0783). Food and Agriculture Organization of the United Nations.FAO. (2019). Faostat statistical database.

Ghandi, A., Navarro, S., & Rivers, D. (2018). On the identification of gross output production

functions.

Goodchild, M., & Zheng, R. (2018). Early assessment of China’s 2015 tobacco tax increase. Bulletin

of the World Health Organization, 96(7), 506-512.Goolsbee, A., & Syverson, C. (2019). Monopsony power in higher education: a tale of two tracks

(Tech. Rep.). National Bureau of Economic Research.Gowrisankaran, G., & Rysman, M. (2012). Dynamics of consumer demand for new durable goods.

Journal of Political Economy, 120(6), 1173–1219.Grieco, P., Pinkse, J., & Slade, M. (2017). Brewed in North America: Mergers, marginal costs and

efficiency. International Journal of Industrial Organization, 59(2018), 24-64.Griliches, Z., & Mairesse, J. (1998). Production functions: The search for identification. In S. Strom

(Ed.), Econometrics and economic theory in the twentieth century (pp. 2413-2415, 2417, 2421).Cambridge University Press.

Gupta, N. (2005). Partial privatization and firm performance. The Journal of Finance, 60(2), 987–1015.

Hall, R. E. (1986). Market structure and macroeconomic fluctuations. Brookings Papers on Economic

Activity, 2(1986), 285-322.Hamilton, J. L. (1994). Joint oligopsony-oligopoly in the US leaf tobacco market, 1924–39. Review

of Industrial Organization, 9(1), 25–39.Harris, J. E. (1998). The price of cigarettes and the profits of cigarette manufacturers with and without

federal intervention, 1997-2006 (Report). American Cancer Society.Hirsch, B., Schank, T., & Schnabel, C. (2010). Differences in labor supply to monopsonistic firms

and the gender pay gap: An empirical analysis using linked employer-employee data fromGermany. Journal of Labor Economics, 28(2), 291–330.

Hsieh, C.-T., & Song, Z. (2015). Grasp the large, let go of the small: The transformation of the state

sector in China (Working Paper No. 21006). NBER.Hu, T., Mao, Z., Ong, M., Tong, E., Tao, M., Jiang, H., . . . Yurekli, A. (2006). China at the crossroads:

the economics of tobacco and health. Tobacco Control, 15(1), 37-41.Imbert, C., Seror, M., Zhang, Y., & Zylberberg, Y. (2018). Migrants and firms: Evidence from China.Lan, L. (2015, 05). Market will now determine tobacco prices. Chinadaily.com.cn.Levinsohn, J., & Petrin, A. (2003). Estimating production functions using inputs to control for

unobservables. Review of Economic Studies, 70(2), 317-341.Li, V., Wang, Q., Xia, N., Tang, S., & Wang, C. (2012). Tobacco crop substitution: pilot effort in

China. American Journal of Public Health, 102(9), 1660-1663.

47

Lu, Y., & Yu, L. (2015). Trade liberalization and markup dispersion: evidence from China’s WTOaccession. American Economic Journal: Applied Economics, 7(4), 221-253.

Manning, A. (2003). Monopsony in motion: Imperfect competition in labor markets. PrincetonUniversity Press.

Matsudaira, J. D. (2014). Monopsony in the low-wage labor market? Evidence from minimum nursestaffing regulations. Review of Economics and Statistics.

Michel, C., & Weiergraeber, S. (2018). Estimating industry conduct in differentiated products mar-

kets.

Miller, N. H., & Weinberg, M. C. (2017). Understanding the price effects of the millercoors jointventure. Econometrica, 85(6), 1763–1791.

Minale, L. (2018). Agricultural productivity shocks, labour reallocation and rural–urban migration inChina. Journal of Economic Geography, 18(4), 795–821.

Morlacco, M. (2017). Market power in input markets: theory and evidence from French manufactur-

ing (Tech. Rep.).Naidu, S., Nyarko, Y., & Wang, S.-Y. (2016). Monopsony power in migrant labor markets. Journal

of Political Economy, 124(6), 1735-1792.Nargis, N., Zheng, R., Xu, S. S., Fong, G. T., Feng, G., Jiang, Y., . . . Hu, X. (2019). Cigarette

affordability in China, 2006–2015: findings from International Tobacco Control China surveys.International journal of environmental research and public health, 16(7), 1205.

Naughton, B. (2007). The Chinese Economy: Transitions and Growth. Cambridge, MA: MIT Press.Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica, 69(2),

307-342.Oaxaca, R. L., & Ransom, M. R. (2010). New market power models and sex differences in pay.

Journal of Labor Economics, 28(2), 267–289.O’Connor, R., Li, Q., Stephens, W. E., Hammond, D., Elton-Marshall, T., Cummings, K. M., . . .

Fong, G. T. (2010). Cigarettes sold in China: Design, emissions and metals. Tobacco Control,19(2), 47-53.

Olley, S., & Pakes, A. (1996). The dynamics of productivity in the telecommunications equipmentindustry. Econometrica, 64(6), 1263-1297.

Peng, Y. (1996). The politics of tobacco: relations between farmers and local governments in China’ssouthwest. The China Journal, 36, 67-82.

Ransom, M. R., & Sims, D. P. (2010). Estimating the firm’s labor supply curve in a “new monopsony”framework: schoolteachers in Missouri. Journal of Labor Economics, 28(2), 331–355.

Ravallion, M., & Chen, S. (2009). China’s (uneven) progress against poverty. In (p. 65-111). Rout-ledge.

Rossi-Hansberg, E., Sarte, P.-D., & Trachter, N. (2018). Diverging trends in national and local

concentration (Tech. Rep.). National Bureau of Economic Research.

48

Scott, P. T. (2013). Dynamic discrete choice estimation of agricultural land use.

Song, F., Zhao, J., & Swinton, S. M. (2011). Switching to perennial energy crops under uncertaintyand costly reversibility. American Journal of Agricultural Economics, 93(3), 768–783.

State Administration of Taxation. (2015). Annual report (Tech. Rep.). State Administration ofTaxation.

State Council of the People’s Republic of China. (1997). Regulation for the implementation of the

law on tobacco monopoly of the People’s Republic of China.

STMA. (2002). Implementation opinions of the state tobacco monopoly administration on the orga-

nizational structure adjustment of cigarette industry enterprises in the tobacco industry.

Sumner, D. A., & Alston, J. M. (1987). Substitutability for farm commodities: the demand forus tobacco in cigarette manufacturing. American Journal of Agricultural Economics, 69(2),258–265.

Syverson, C. (2004). Market structure and productivity: A concrete example. Journal of Political

Economy, 112(6), 1181–1222.Syverson, C. (2019). Macroeconomics and market power: Context, implications, and open questions.

Journal of Economic Perspectives, 33(3), 23–43.Talhout, R., Richter, P. A., Stepanov, I., Watson, C. V., & Watson, C. H. (2018). Cigarette design

features: effects on emission levels, user perception, and behavior. Tobacco regulatory science,4(1), 592–604.

Tortarolo, D., & Zarate, R. (2018). Measuring imperfect competition in product and labor markets

(Mimeo). UC Berkeley.United Nations. (2019). Un comtrade database. (http://comtrade.un.org/)U.S. Census Bureau. (1997a). 1997 economic census of manufacturing.

U.S. Census Bureau. (1997b). Cigarette manufacturing (1997 Economic Census - Industry Series).US Census Bureau.

Van Biesebroeck, J. (2007). The robustness of productivity estimates. Journal of Industrial Eco-

nomics, 55(3), 529-569.Wang, J. (2013). State-Market Interaction in China’s Reform Era: Local State Competition and

Global Market-building in the Tobacco Industry. Routledge.Wollmann, T. (2019). Getting away with merger: The case of dialysis clinics in the United States.

(Working Paper)World Economic Forum. (2019). World Economic Forum Global Competitiveness Index.

Yang, D. T. (1999). Urban-biased policies and rising income inequality in China. American Economic

Review: Papers and Proceedings, 89(2), 306-310.Zheng, R., Wang, Y., Hua, X., & Marquez, P. V. (2016). Cigarette affordability in China, 2001-2016.

World Bank.

49

Figure 1: Tobacco industry structure

(a) Value chain

Farmers (20M)

Cigarette manufacturers (350 → 150)

Wholesaler (monopolist)

Retailers

Consumers

Tobacco leaf

Cigarettes

(b) Farming

(c) Drying

(d) Manufacturing

Notes: Figure (a) gives a schematic overview of the consecutive actors in the cigarette value chainin China. “CNTTC” stands for Chinese National Tobacco Trade Company, and is the wholesalingarm of the CNTC/STMA. This is a government-controlled monopolist. Figures (b)-(d) show imagesof the production process steps. Copyrights go to (a) ©Hu(2018), (b) ©news.cn, and (c) ©GettyImages

50

Figure 2: Cigarette manufacturing locations

(a) 1999

(b) 2006

Notes: These maps depict all counties with at least one cigarette manufacturing firm in 1999 and2006. In counties with at least one cigarette manufacturer, there were on average 1.24 firms.

51

Figure 3: Market structure

(a) All firms

.4

.5

.6

.7

.8

Con

cent

ratio

n ra

tio

0

50

100

150

200

250

300

350

# Fi

rms

1999 2000 2001 2002 2003 2004 2005 2006

# Manufacturers CR1CR2

(b) Firms below vs. above size threshold

0

20

40

60

80

100

# Fi

rms

1999 2000 2001 2002 2003 2004 2005 2006

Q<100K cases Q>100K cases

Notes: Panel (a) shows the evolution of the total number of cigarette manufacturers in China (leftaxis) and the combined provincial market shares of the two biggest firms per province (right axis).Panel (b) breaks this evolution down into firms below and above the exit threshold of 100,000 casesper year. This graph excludes firms for which quantities are unknown, which is why the total numberof firms in panel (b) is lower compared to panel (a).

52

Figure 4: Factor revenue shares

(a) Aggregate

0

.1

.2

.3

.4

Expe

nditu

re/R

even

ue

1999 2000 2001 2002 2003 2004 2005 2006Year

Leaf Labor

(b) By treatment (average)

6

8

10

12

14

Leaf

/Lab

or C

osts

1999 2000 2001 2002 2003 2004 2005 2006

No firms with Q<100K One or more firms with Q<100K

(c) By treatment, weighted average

6

8

10

12

14

Leaf

/Lab

or C

osts

1999 2000 2001 2002 2003 2004 2005 2006Year

No firms with Q<100K One or more firms with Q<100K

(d) By treatment, median

6

8

10

12Le

af/L

abor

Cos

ts

1999 2000 2001 2002 2003 2004 2005 2006

No firms with Q<100K One or more firms with Q<100K

Notes: Panel (a) plots the evolution of the total wage bill and total intermediate input expenditureover industry revenue. Panels (b) to (d) plot the ratio of labor expenditure over intermediate inputexpenditure over time for two groups of firms. The treatment group includes firms with at least onecompetitors below the size threshold in 2002, while the control group are firms without any of thesecompetitors in 2002. In panel (b), I calculate the unweighted average for the control and treatmentgroups. In panel (c), I weight by labor expenditure. In panel (d), I calculate the median. Marketdefinitions are at the prefecture level.

53

Figure 5: Markups and markdowns

(a) Leontief model

0

.5

1

1.5

Kern

el d

ensi

ty

1 2 3 4 5 6

Markdown Markup

(b) Cobb-Douglas model

0

.5

1

1.5

2

Kern

el d

ensi

ty

0 5 10 15

Markdown Markup

Notes: Panel (a) plots markdowns and markups using the baseline model in which tobacco leaf isnot substitutable. Panel (b) does the same using the Cobb-Douglas model with substitutable tobaccoleaf. All distributions are censored at 5th and 95th percentiles.

54

Figure 6: No consolidation counterfactual

-.5

0

.5

1

Rel

ativ

e ch

ange

with

200

3

1999 2000 2001 2002 2003 2004 2005 2006Year

Manufacturing wage Leaf priceLeaf price CF

Notes: The solid lines plot change in average manufacturing wages and leaf prices compared to2003 (normalization: 2003=0). The dashed line plots the counterfactual leaf price evolution in thecounterfactual scenario in which the exit thresholds were not enforced.

55

Table 1: Selected summary statistics

Variable Mean Std. Dev. N

Revenue (million $) 104.6 195.23 1109Quantity (million cases) 0.34 0.42 1109Price per case ($) 1623.09 14118.72 1109Profit (million $) 11.73 42.52 1109Wage bill (million $) 3.46 6.06 1109Employees (thousands) 1.19 1.06 1109Material expenditure (million $) 35.59 51.16 1109Capital stock (million $) 47.71 71.14 1109Export dummy 0.22 0.42 1109Export share of revenue 0.01 0.05 1109County population (millions) 71.67 49.52 792Leaf content per cigarette (mg) 681.47 31.07 185Filter density (mg/ml) 112.8 3.68 185

Notes: A case contains 50,000 cigarette sticks. Reported prices arefactory-gate prices. All monetary variables are denoted in 2006 USdollars.

56

Table 2: Reduced-form evidence: consolidation and prices

Panel (a): Province-level markets log(Wage) log(Leaf price) log(Cigarette price)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.160 (0.086) -0.467 (0.116) -0.391 (0.091)

R-squared 0.800 0.862 0.863Observations 1,091 1,091 1,091

Panel (b): Prefecture-level markets log(Wage) log(Leaf price) log(Cigarette price)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.128 (0.082) -0.466 (0.120) -0.257 (0.100)

R-squared 0.800 0.865 0.863Observations 1,091 1,091 1,091

Panel (c): County-level markets log(Wage) log(Leaf price) log(Cigarette price)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.056 (0.117) -0.747 (0.170) -0.349 (0.123)

R-squared 0.800 0.870 0.864Observations 1,091 1,091 1,091

Panel (d): Pre-trends log(Wage) log(Leaf price) log(Cigarette price)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * Year -0.031 (0.090) -0.189 (0.156) -0.145 (0.137)

R-squared 0.031 0.021 0.020Observations 764 764 764

Notes: I estimate how labor wages, leaf prices and factory-gate cigarette prices change differently between firmswith and without any competitors below the exit threshold in 2002. Panel (a) defines the input market at theprovince-level, panel (b) at the prefecture-level and panel (c) at the county-level. Controls include firm fixedeffects, export dummy, ownership type dummies and a linear time trend. Panel (d) estimates the pre-trends forall three outcomes in the period 1999-2003, with treatment effects being defined at the province-level.

57

Table 3: Structural model estimates

Panel (a): Output elasticities ACF IVEstimate: SE: Estimate: SE:

Labor 0.320 (0.226) 0.404 (0.096)

Capital 0.761 (0.140) 0.699 (0.107)

Returns to scale 1.081 (0.191) 1.104 (0.079)

R-squared 0.989 0.8871st stage F-statistic – 46.79Observations 823 1,108

Panel (b): Leaf supply semi-elasticity OLS IVEstimate: SE: Estimate: SE:

Leaf price -0.172 (0.097) 3.671 (0.862)

R-squared 0.863 0.1191st stage F-statistic – 79.14Observations 1,091 1,091

Panel (c): Labor supply semi-elasticity OLS IVEstimate: SE: Estimate: SE:

Labor -0.0001 (0.001) -0.029 (0.068)

R-squared 0.904 0.7931st stage F-statistic – 21.89Observations 1,091 1,091

Notes: Panel (a) reports the output elasticities which are obtained from estimating the pro-duction function. Using ACF reduces the sample size as the lagged variables need to beobserved. Panels (b) and (c) report the estimated price semi-elasticities of input supply, bothfor tobacco leaf and labor. The left-hand side variable is the log province-level market shareminus the log outside option market share. The endogenous right-hand side variable is the leafprice for one pack of cigarettes in 1000 RMB. Manufacturing TFP from the Leontief modelis used as instrument. I control for prefecture dummies, ownership dummies, and cigaretteprices. Unit wages are also controlled for in the leaf supply estimation model. Bootstrappedstandard errors are in the parentheses.

58

Table 4: Markups and markdowns

Panel (a): Leontief model Markdown MarkupEstimate S.E. Estimate S.E.

Mean 2.869 (0.433) 1.062 (0.092)

Median 2.235 (0.286) 0.923 (0.086)

Observations 1,091 1,091

Panel (b): Cobb-Douglas model Markdown MarkupEstimate S.E. Estimate S.E.

Mean 0.768 (1.510) 5.629 (8.127)

Median 0.371 (0.621) 3.440 (5.946)

Observations 1,091 1,091

Panel (c): Correlations log(Markdown) log(Markup)Estimate S.E. Estimate S.E.

log(Output) 0.113 (0.068) 0.064 (0.036)

1(SOE) 1.266 (0.221) -0.064 (0.119)

log(Unemployment rate) 0.233 (0.106) -0.112 (0.057)

log(Migration rate) -0.304 (0.116) 0.129 (0.053)

log(Tax rate) 0.285 (0.054) -0.055 (0.045)

R-squared 0.577 0.862Observations 751 751

Notes: Panel (a) shows key moments for the markdown and markup distribution us-ing the baseline Leontief model in which tobacco leaf cannot be substituted for laboror capital. Panel (b) does the same for the model that is Cobb-Douglas in leaf, laborand capital. Panel (c) shows correlations of the markup and markdown distributionfrom the Leontief model with selected firm and county characteristics. Province-dummies are controlled for. Bootstrapped standard errors are in parentheses.

59

Table 5: Consolidation treatment effects

Panel (a): Leontief model log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.314 (0.065) -0.265 (0.065) 0.075 (0.082)

1(year≥2003) -0.320 (0.071) 0.334 (0.075) -0.159 (0.091)

R-squared 0.811 0.764 0.867Observations 1,091 1,091 1,091

Panel (b): Cobb-Douglas model log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.223 (0.092) -0.147 (0.068) 0.298 (0.084)

1(year≥2003) -0.083 (0.100) 0.088 (0.078) -0.339 (0.092)

R-squared 0.731 0.758 0.871Observations 1,091 1,091 1,091

Panel (c): Pre-trend log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment*Year 0.089 (0.084) -0.077 (0.093) -0.054 (0.042)

Treatment -178.452 (167.755) 154.931 (185.117) 107.538 (84.231)

Year -0.046 (0.082) 0.095 (0.090) 0.042 (0.030)

R-squared 0.316 0.095 0.404Observations 756 756 756

Notes: Panel (a) estimates the differences-in-differences model using the Leontief specification and province-level input market definitions. Controls include firm fixed effects, a linear time trend, ownership type dummies andexport dummies. Panel (b) estimates the same regression, but uses the Cobb-Douglas model in labor, capital andtobacco leaf instead. I use the same sample as in panel (a), even though the Cobb-Douglas markups and markdownsare available for a larger sample as they do not require observable quantities. Panel (c) estimates the pre-trends onthe period 1999-2002, and uses the Leontief model estimates. Bootstrapped standard errors are in parentheses.

60

Table 6: Heterogeneous treatment effects

Panel (a): Firm characteristics log(Markdown) log(Markdown)Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.808 (0.756) 4.925 (1.134)

Treat*1(year≥ 2003)*#firms under 100K 1.379 (0.607)

Treat*1(year≥ 2003)*firm output -0.364 (0.096)

R-squared 0.338 0.470Observations 779 771

Panel (b): County characteristics log(Markdown) log(Markdown)Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 2.326 (0.813) 3.948 (1.017)

Treat*1(year≥ 2003)*log(unemployment rate) 0.457 (0.215)

Treat*1(year≥ 2003)*log(migration rate) 0.832 (0.268)

R-squared 0.336 0.327Observations 779 771

Notes: Panel (a) interacts the consolidation treatment effects with the number of competitors pro-ducing below the exit threshold before 2003 and with the firm’s output level. Panel (b) interactsthe treatment effects with the county-level unemployment and migration rates. The migration rate isdefined as the share of the county population who are born in another province.

61

Table 7: Consolidation and margins along the chain

Panel (a): Share of industry surplus Farming Manufacturing Wholesale+Retail

0.286 0.352 0.362

Panel (b): Log(GPM∗) of: Farmers Manufacturers WholesalersEstimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.498 (0.101) 0.091 (0.055) 0.273 (0.089)

Observations 1,091 1,091 697R-squared 0.846 0.684 0.873

Notes: ∗Gross Profit Margin. Panel (a) shows the division of industry surplus between farmers, manufacturersand wholesalers. Province-level market definitions are used. Panel (b) estimates how the gross profit marginsof farmers, manufacturers and wholesalers changed due to the consolidation. Linearly fitted retail prices areused in 1999 and 2000. The three effects do not need to add up to zero as consumer surplus changed as well.Province-level market definitions are used.

62

Table 8: Scale economies

Panel (a): Province level log(Fixed costs) log(Capital stock)Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.310 (0.356) -0.051 (0.309)

R-squared 0.295 0.236Observations 213 213

Panel (b): Prefecture level log(Fixed costs) log(Capital stock)Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.269 (0.241) 0.043 (0.192)

R-squared 0.154 0.079Observations 743 763

Panel (c): County level log(Fixed costs) log(Capital stock)Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) -0.211 (0.257) 0.151 (0.210)

R-squared 0.266 0.253Observations 933 993

Notes: I estimate how total fixed costs and the capital stock at the input marketlevel were affected by enforcing the exit thresholds. I define input markets atthe province, prefecture and county level in panels (a), (b) and (c). The numberof observations is lower for fixed costs than for the capital stock as total inferredfixed costs are negative for some markets, which are omitted due to taking logs.

63

Appendices

A Data appendix

A.1 Production and cost data

As the dataset was already cleaned in Brandt et al. (2012), who checked consistency of firm identifiers,for instance, not much more cleaning is required. I do remove outliers in cigarette and leaf pricesby censoring at the 1st and 99th percentiles. I also delete observations with negative intermediateinput expenditure. As a result, I retain 2,025 observations, covering 470 firms. When defining inputmarkets and treatment effects and when estimating markdowns, I include all firms in this industry, notjust cigarette producers, in order to cover all tobacco leaf buyers. In the remainder of the analysis,I restrict the analysis to cigarette manufacturers only, which account for 95% of industry revenue,and also only focus on firms for which quantities are observed. This reduces the dataset to 1,091observations.

A.2 Quantity data

Quantities are given at the product-firm-month level during these years by the NBS, and as the firmidentifiers are the same as in the ASIF dataset, both can be merged. I only keep product codes that aremeasured in numbers and aggregate to the yearly level. As firms usually produce just one product,cigarettes, firm-level prices can be inferred simply by dividing firm revenue by the total number ofunits produced per year. Quantities are observed for just about half of the firms, which reduces thesample size to 1260. Units are defined as cigarette cases, each of which contain 50,000 cigarettes(Fang et al., 2017). From 2004 onwards, the case definition seems to change. Fortunately, I observeboth current and lagged quantities in each month, and by comparing both I am able to scale the post-2004 quantities in order to make them consistent with the pre-2004 observations. As the treatmentvariables are defined based on quantity units in 2002, this does not change cross-sectional variation inmonetary variables. More details about this dataset are in Lu and Yu (2015).

I deflate all monetary variables using the relevant deflators but as I study just one industry this onlyaffects the time-series, and not the cross-sectional, variation in the observed variables.

64

A.3 County data

I retrieve county-level data from the 2000 population census through the Harvard Dataverse. Thepopulation census contains many variables, of which I use the total county population, the unem-ployed population, and the number of immigrants per county.

A.4 Product characteristics

I obtain brand-level cigarette characteristics from O’Connor et al. (2010) for a subset of firms in 2009,such as the leaf content per cigarette and other characteristics which affect the smoking experience(Talhout, Richter, Stepanov, Watson, & Watson, 2018). This dataset is observed only for 13% of theobservations, but covers 29% of observations by revenue. Again, I only use this data in an extension.I link the brands in O’Connor et al. (2010) to the firms in the dataset. As I do not observe a decom-position of firm sales into brands, I have to aggregate from the brand to the firm-level, and therefore Isimply calculate average product characteristics across brands.

A.5 Sample size

After cleaning the data, such as dropping observations with negative input expenditures, I retain 2,025observations in total, 1,091 of which contain observed quantities. The main regressions which requirequantity data will hence have a sample size of 1,091.

B Derivations

B.1 Markups with endogenous input prices and non-substitutable inputs

I start by deriving the markup expression in equation (6a). I start with the cost minimization problemin equation (5). Firms choose input prices in order to minimize per-period costs. The shadow priceλft is the marginal cost of increasing the input price by one unit. The markup µ is, as always, definedas the ratio of the product price over this marginal cost:

µft ≡Pftλft

65

Solving the first order cost minimization condition gives the following expression for λ:

λft = WLft

∂Lft∂Qft

+WMftMft

Qft

ψMft

Substituting the revenue shares αVft ≡VftW

Vft

PftQft

for V ∈ L,M and βLft ≡∂Qft

∂Lft

LftQft

gives:

λft =αLftPft

βL+ αMftPftψ

Mft

Finally, dividing prices by marginal costs yields equation (6a).

B.2 Markdown interpretation of input supply elasticity

Next, I derive the markdown formula in equation (7). In contrast to the previous paragraph, I nowassume that firms choose the leaf price to maximize profits rather than minimize costs (this will, ofcourse, give the same result). The first order condition gives:

∂PftQft

∂Mft

∂Mft

∂WMft

−Mft −WMft

∂Mft

∂WMft

−WLft

∂Lft∂Qft

∂Qft

∂Mft

∂Mft

∂WMft

= 0

Dividing by ∂Mft

∂WMft

and substituting the expression for ψM leads to equation (7).

B.3 Demand for substitutable inputs with endogenous prices

Firms choose the optimal leaf price WMft∗ that minimizes their costs, as shown in equation (5). This

leaf price corresponds to an optimal output level Q∗ft, which is a function of the the markdown ψM ,all input prices, Hicks-neutral productivity, and the leaf requirement βM .

Conditional on this optimal output level, firms choose the mix between labor and capital. Denotingthe marginal cost of labor and capital as λHft, and the interest rate as WK

ft , the choices for labor andcapital are given by the following cost minimization problem:

minLft,Kft

WLftLft +WKKft − λHft

(H(Lft, Kft)−Q∗ft

)Assuming a Cobb-Douglas function H(L,K), solving the first order conditions results in labor

demand l(kft, wLft, wMft , β

L, βK ,WK , ωft, µft, ψMft ).

The optimal amount of labor used depends on the leaf price markdown. The reason for this is that

66

if leaf prices react more to the leaf quantity used, firms choose a lower equilibrium output level, andhence also use less labor.

B.4 Input market equilibrium in oligopsony

The conditions for the leaf market to be in equilibrium are needed to solve equilibrium leaf prices ineach iteration of the Monte Carlo simulation. Let product prices P be normalized to one and let firmsf simultaneously choose input prices WM

ft every period by solving the same static cost minimizationproblem as before:

WMft = arg min

(WLftLft +WM

ftMft −Q(.))

⇔ WLft

∂Lft∂Mft

∂Mft

∂WMft

+WMft

∂Mft

∂WMft

+Mft −∂Qft

∂Lft

∂Lft∂Mft

∂Mft

∂WMft

= 0

⇔Mft +∂Mft

∂WMft

(WMft + βMft

WLft

∂Qft∂Lft

− βMft)

= 0

Dividing by market size M and defining Sft ≡ Mft

M, this gives the following equilibrium condition,

which is very similar to those in oligopoly demand models:

Sft +∂Sft∂WM

ft

(WMft + βMft

WLft

∂Qft∂Lft

− βMft)

= 0 (15)

B.5 Different objective functions

Assumption (1) stated that firms decide on their input prices to minimize per-period variable costs.SOEs could, however, have different objectives, such as achieving ‘social stability’ through high andcountercyclical employment, as argued by Li et al. (2012). There are two ways in which such sizeobjectives can enter the firm’s minimization problem.

Output size objective

Suppose first that firms value being large and are willing to sacrifice some profits to achieve this. Suchan objective enters the shadow price λft. Let this shadow price in equation (5) be denoted λft =

λftτft

.Firms with a larger preference for producing a lot have a larger parameter τft. The markup µft is nowgiven by:

µft = τftPftλft

67

If firms value being large rather than profitable, the true markup will hence be larger than theestimated markup. The reason for this is that the cost minimization model infers large input usage asan indication of low markups, while in reality, this is due to a preference towards large size.

The same holds for the markdown ψMft . If firms value a large size, they will set higher input prices.Through the markdown expression, the model interprets this as evidence for low monopsony power.In reality, though, this merely reflects size preferences.

ψMft = γWWMft (1− Sft)−1 + 1

Input size objective

Now suppose that firms specifically want to employ a lot of manufacturing workers, but do not havesuch preferences for farming employment (or the other way around). In this case, the true input priceWLft is different from the measured input price. If firms value employing many workers, the implicit

wage is lower than the observed wage, so τLft < 1:

WLft = WL

ftτLft

As firms do not choose labor and tobacco leaf separately, this has the same effects on markup andmarkdown estimates as a different shadow price λ. The derivation in appendix B shows that marginalcosts are linear in both input prices.

Remedies

Throughout the text, I always allow for different objective functions across ownership types, becauseI systematically add ownership dummies to all regressions. These never change the results much, asthere is not much variation in ownership: most firms are SOEs anyway. I do not allow for differentobjective functions within ownership classes, as all SOEs are, for instance, assumed to have the sameobjective function.

68

C Robustness checks

C.1 Difference-in-differences robustness checks

Continuous treatment measures

Throughout the main text, a dummy variable was used to indicate the treatment groups, namely thepresence of firms below the exit threshold in a market before 2003. I re-estimate the treatment modelusing different treatment measures. First, I use the share of firms in a market producing below thethreshold in 2002 as a treatment measure. Second, I weight these firms by employment, rather thantaking unweighted averages. Besides the treatment indicator definition, I keep all model specificationsfixed. The results are in panel (a) of table A8. The interpretation of the estimates is as follows: inprovinces in which the share of firms below the threshold was 10 percentage points higher, markdownswere 8.3% higher. The results are consistent with the baseline regression: increasing markdowns inconsolidated markets, decreased markups and no strong evidence for productivity gains.

Firm and year fixed effects

In the baseline model, firm fixed effects were included as controls, while year fixed effects were not.A linear time trend was used instead. In panel (b) of table A8, I use two different specifications. First,I no longer control for firm dummies. The coefficients are similar to the model with firm fixed effects,except that the drop in markups is less pronounced. Next, I include both firm and year fixed effects.This does not change the estimates much.

Different moments

The baseline difference-in-differences model was estimated using regular unweighted OLS. In thefirst part of panel (c) of table A8, I use a quantile regression instead. I report the treatment effects onthe median of each outcome variable. The estimates are now smaller compared to when averages areused, and are less dependent on the market definitions used. The broad trend of rising markdowns andfalling markups is still present in all specifications, though. Finally, I weight observations by theiremployment size. The estimated coefficients are very similar to the baseline regression.

69

Self-selection above exit threshold

The exit thresholds of 100,000 cases per year were based on production from 2002. If firms knewthese exit thresholds beforehand, there should be ‘bunching’ of firms just above this exit threshold.Figure A1 plots the distribution of the number of cases between 2000 and 2004. Up to 2003, mostfirms are around the exit threshold of 100,000 cases per year, and there is no bunching just above thisthreshold.

Dropping firms under the 100K threshold

In table A11, I re-estimate the difference-in-differences regression when dropping the firms under theproduction threshold which survived after 2002. The results are similar to the main specifications.

Narrowing the treatment time window

As was shown in figure 3, the number of firms producing less than 100,000 cases a year fell rela-tively faster mainly between 2002 and 2004. After 2004, there was also significant exit among firmsproducing more than 100,000 cases a year.

I narrow down the timeframe to the periods 1999-2005 and 1999-2004, which means that the treat-ment effects are estimated over just two and one years respectively, in table A10. The effects becomea bit smaller when narrowing the time frame, but are still in the same order of magnitude as the orginalresults, and signficant.

Announcement effects

The industry consolidation and specific size thresholds were announced by the STMA on May 2,2002. It may be, however, that firms were anticipating the reform and decreased their leaf pricesalready before the policy was implemented. In table A9, I test for such announcement effects. Ire-estimate the difference-in-differences model from equation (1) with markdowns, markups and pro-ductivity as the outcome variables. In the main specification, the start of the policy was defined to be2003. I also re-estimate the difference-in-differences model using 2001 and 2002 as treatment startingyears. I use province-, prefecture- and county-level market definitions in panels (a), (b) and (c).

At both the province- and county-level, markdowns decreased significantly already in 2002, whichis consistent with announcement effects as the policy was announced in that year. Markups seemto have decreased, however, already before 2002. These markups reflect pricing behavior by thegovernment-owned wholesaler. As the STMA, which ordered the consolidation, and the CNTC,

70

which operates the wholesaling monopoly, are heavily intertwined, it is possible that they adjustedtheir prices in expectation to the policy change. The modest rise in productivity after 2003 in thetreatment group was, finally, not present before the treatment started.

C.2 Translog production function

Throughout the main text, I used a Cobb-Douglas specification for the labor-capital term H(.) inthe production function. As the elasticity of substitution estimate between labor and capital was notsignificantly different from one, this seems to be the correct production function. Nevertheless, I alsouse a translog specification for H(.) as a robustness check. The corresponding functional form of h(.)in logarithms is given by:

h(Lft, Kft) = βLlft + βKkft + βLK lftkft + β2Ll2ft + 2β2Kk2ft

The moment conditions to estimate this translog production are given by:

Eξft(β

L, βK , βLK , βL2, βK2)

lft−1

kft

lft−1kft

l2ft−1

k2ft

= 0

Table A3 compares the estimates for the main coefficients of interest and the markdown and markupaverages between the Cobb-Douglas and Translog model. The estimates are very similar and neversignficantly differ from each other.

C.3 Using cost shares as output elasticities

Suppose there are constant returns to scale, exogenous input prices for labor and capital, and a Cobb-Douglas function for H(.) in the production function, equation (3). The output elasticities of laborand capital are then equal to their cost shares. Denoting capital investment as Ift, this gives:

βLft =

WLftLft

WLftLft+Ift

βKft =Ift

WLftLft+Ift

The output elasticities, markup and markdown estimates and consolidation treatment effects usingthis approach are in table A4. They are very similar to the baseline estimates.

71

C.4 Alternative labor supply specifications

Log-logs and inverse supply

The baseline supply model was a log on levels labor supply model, in equation (8). Alternatively, Iestimate the following log-on-logs model:

sft − s0t = γW log(WMft ) + γZZft + ζft

I can also estimate an inverse supply model, as in Goolsbee and Syverson (2019):

log(WMft ) = γW log(sft − s0t) + γZZft + ζft

Nested logit

It is possible to allow for more flexible substitution patterns in supplier choices by using a nestedlogit model. Each farmer j now chooses a manufacturer f within nest g in market i. The set ofmanufacturers in nest g in market i at time t is denoted Fgit. The error structure in the utility functionnow differs, with utility being parametrized as:

Ujft = γWWMft + γZZft + ζjt︸ ︷︷ ︸

δft

+(1− σ)νjft

The preference shocks νjft still follow a type-I extreme value distribution. Following S. T. Berry(1994), the input market share is given by:

Sft =exp(

δft1−σ )

Dσgt[∑

gD1−σgt ]

with Dgt ≡∑

f∈Fgitexp(

δft1−σ ) The markdown is now expressed as:

ψMft ≡( ∂Sft∂WM

ft

WMft

Sft

)−1

+ 1 =

(γWWM

ft

( 1

1− σ −σ

1− σSfgt − Sft))−1

+ 1

I use two types of nests. In the first specification, I define three nests, according to the manufac-turer’s ownership types. I divide manufacturers into (i) SOEs, (ii) private firms and collectives, and(iii) foreign firms. Secondly, I define prefectural markets as nests of province-level markets.

72

Results

The estimates for these alternative labor supply specifications are in table A6. In panel (a), I reportthe log-on-logs estimates. These are consistent with the baseline specification: the leaf supply curveis upward-sloping, the labor supply curve flat. Panel (b) reports the inverse log-on-log specification.If leaf demand increases by 10%, leaf prices increase by 1.5%. For labor, the coefficient is negativebut insignificant. Panels (c) and (d) contain the estimates for the nested logit models. The nestingelasticitiy between different prefectural sub-markets is 0.621. When defining the nests as ownershiptypes, the elasticity is 0.545. Both elasticity estimates are significantly below one and above zero,meaning that the nests are neither perfect complements nor substitutes.

The average and median markup and markdown estimates from both nested logit models are intable A7. Markdowns are on average higher in the nested logit models: 2.85 with the ownershipnests and 3.05 with the prefectural nests, compared to 2.63 in the standard logit model. Markupsare - mechanically - lower, at 1.05 and 0.96. The treatment effects of the consolidation are verysimilar to those in the baseline model. Finally, I compare the markup and markdown distributionsfrom both nested logit models with the baseline logit model in figure A2. The markdown distributionbecomes flatter when allowing for more flexible substitution patterns, which is logical: more variationin markdowns is allowed for, both across firms and over time.

C.5 Export participation

The Chinese economy globalized rapidly throughout the 1990s and 2000s, and the accession to theWTO in 2003 affected total factor productivity and markup growth (Brandt, Van Biesebroeck, Wang,& Zhang, 2017; Li et al., 2012). As was already argued, the tobacco industry remained largely domes-tic: less than 1% of total industry revenue came from exports. For the 16% of firms who do export,exports represent merely 6% of their revenues on average. In table A20, I test whether exportingbehavior changed through the consolidation, and conclude that it did not.

D Monte Carlo simulations

D.1 Motivation

In this section, I simulate a production and cost dataset with oligopsonistic input markets. This servestwo objectives. First, I compare consistency and efficiency of existing production function estimationapproaches with the IV approach proposed in this paper, under different models of input market

73

competition. Second, I use this data to show that the input supply curve can be consistently estimatedusing productivity shocks of buyers as instruments. The estimation strategy in the paper consists ofthree broad steps:

(i) Estimate the production function, equation (10a)→ retrieve (ωft, βft)

(ii) Use productivity residuals ωft to estimate markdowns by using re the input supply function andmarkdowns→ retrieve (ψMft )

(iii) Retrieve markups using βft from (i) and ψMft from (ii).

D.2 Data generating process

I create a panel of 200 local intermediate input markets with exactly three firms in each market. Allfirms are observed during 3 periods and there is no entry or exit. As a result, the dataset contains 1800observations. I impose a standard logit model for intermediate input supply, as in the model section.I simulate the dataset using 50 iterations.

Simulation procedure

In each iteration loop, the simulation procedure is as follows:

1. Draw ωf1, ξf1, ζf1, Kf1

2. Find equilibrium wages WMf1 using conditions in equation (15)

3. Calculate optimal investment If1

4. Use transition equations to get ωf2, xf2, Kf2

5. Repeat steps (2)-(4) until final period

I start by drawing initial productivity ω, productivity shocks ξ, firm characteristics ζ (which entersupplier utility) and capitalK from their respective distributions, all in the first year. Next, I search forequilibrium input prices and intermediate input quantities for all firms in order to have an equilibriumin all input markets. Based on these equilibrium input prices, firms decide on investment, and optimalinvestment is calculated. In a fourth step, the capital stock is updated for the next year, using thecapital accumulation equation and productivity is updated in the next period using the equation ofmotion for productivity. In the next period, the draws and equilibrium calculation are repeated. Dueto the upward-sloping intermediate supply curve, analytical expressions for labor usage no longer

74

exist (in contrast with, for instance, (Van Biesebroeck, 2007)). Equilibrium wages and employmentlevels are jointly obtained by solving the equilibrium conditions in equation (15) for each market ineach year.

Parametrization

The parametrization of all coefficients and distributions are in table A15. I let the output elasticity oflabor and capital be 0.4 and 0.6. I use a lognormal distribution for TFP, with a serial correlation of 0.7and a cross-sectional standard deviation of 0.5. Input market sizes are distributed uniformly between0.5 and 1.5, meaning that the largest markets are three times larger than the smallest ones. Workingconditions are drawn from a uniform distribution between 0 and 3. Valuation for wages and workingconditions is 2.5 and 0.5 respectively, and in the second DGP there is a random coefficient on wageswhich varies 15% compared to the average wage valuation. Investment costs are normalized to oneand the annual depreciation rate is 10%.

[Table A15 here]

D.3 Results

The estimated output elasticities using the simulated dataset are in panel (a) of table A16. The OLSestimates are, as usual, biased due to simultaneity. When input prices and market shares do not fea-ture in the first stage regression of ACF, the ACF estimates are as biased as OLS, which confirms thenon-identification result in the theoretical model. When properly controlling for both market sharesand input prices, in column 3, ACF delivers consistent estimates.

[Table A16 here]

Secondly, I use the productivity residuals as instrument for input prices in the supply estimation.In the simulated model, the wage valuation coefficient is 2.5. Panel (c) in table A16 compares theOLS and IV estimates of equation (8). The results are very similar to the real estimates in the paper:The OLS estimate is negative, while the IV estimate is positive and close to the truth, at 2.25. Thetrue value of 2.50 lies within the 95% confidence interval of the estimates. The first stage F-statisticis around 100, meaning that TFP is a strong instrument for endogenous input prices.

75

Random coefficients

In the model, it was assumed that input supply follows a standard logit model. This implies thatinput suppliers all value input prices in the same way. It could be, however, that there is unobservedheterogeneity in supplier preferences for input prices and other firm characteristics. In the case oflabor, for instance, it seems likely that some workers prefer higher wages while others value workingconditions or job security more. While estimating such a random coefficients input supply function ispossible using mixed logit models, such as S. Berry et al. (1995), it poses problems for identifying theproduction function. Let the input supply elasticities be distributed across suppliers, and hence acrossfirms with mean γW and σγ:

γWft ∼ Γ(γW , σγ)

In this case, markdown variation across firms is not captured by observed input prices and marketshares alone. Including these in the input demand function hence does not suffice in solving theproblem of serially correlated unobservables which prevent inverting unobserved scalar productivity.Panel (b) in table A16, in the appendix, shows the simulated results for both production functionestimators when input supply follows a random coefficients logit model. The estimates when usingAckerberg et al. (2015) are almost as biased as the OLS results, even when including market sharesand wages in the first stage regression.

Solving this problem is beyond the scope of this paper: for agricultural spot markets, it seemsreasonable to assume that farmers all value input prices in the same way. For labor markets, thisseems more problematic. The solutions to this challenge are reserved for further work.

76

Figure A1: Quantity distribution and bunching

2.00

0e-0

64.

000e

-06

6.00

0e-0

6Pr

obab

ility

dens

ity

50000 100000 150000 200000 250000 300000Annual output in cases

2001 20022003 2004

Notes: This graph plots the distribution of the number of cigarette cases in 2001, 2002 and 2003.There is no evidence for ‘bunching’ just above the exit threshold of 100,000 cases per year.

77

Figure A2: Markups and markdowns: nested logit

(a) Markdowns

0

.1

.2

.3

.4

.5

0 2 4 6 8Markdown

Standard logit Geographical nestsOwnership nests

(b) Markups

0

.5

1

1.5

0 .5 1 1.5 2Markdown

Standard logit Geographical nestsOwnership nests

Notes: Panel (a) plots markdowns using the baseline logit model and both nested logit models, withnests being defined as ownership types and as prefectural nests. Panel (b) does the same for markups.All distributions are censored at 5th and 95th percentiles.

78

Table A1: Treatment and control groups before policy

Panel (a) Treatment group size % Firms % Revenue % Output

Firms producing less than 100K cases 50.4 8.1 4.3

Treatment (province) 86.1 89.2 87.7

Treatment (prefecture) 38.8 38.3 40.4

Treatment (county) 14.9 11.7 15.9

Panel (b) Observable characteristics Cigarette price∗ Leaf price∗∗ Quantity∗∗∗

Mean S.D. Mean S.D. Mean S.D.

Control group (province) 1,384 (613) 546 (177) 568 (433)

Treatment group (province) 1,412 (688) 602 (249) 491 (279)

Control group (prefecture) 1,448 (733) 626 (251) 575 (232)

Treatment group (prefecture) 1,346 (5 67) 559 (120) 626 (251)

Control group (county) 1,447 (711) 705 (243) 601 (255)

Treatment group (county) 1,186 (428) 535 (152) 559 (120)

Notes: ∗Price per case of 50,000 cigarettes in RMB, ∗∗Leaf price per case, ∗∗∗Annual output inthousands of cases. All comparisons are on the period 1999-2002.

79

Table A2: Treatment effects and HHIs

log(Materials HHI) log(Labor HHI)Panel (a): Province level Estimate: SE: Estimate: SE:

Treatment*1(year≥2003) 0.269 (0.081) 0.293 (0.086)

R-squared 0.332 0.400

Panel (b): Prefecture level Estimate: SE: Estimate: SE:

Treatment*1(year≥2003) 0.181 (0.044) 0.186 (0.044)

R-squared 0.172 0.156

Panel (c): County level Estimate: SE: Estimate: SE:

Treatment*1(year≥2003) 0.122 (0.031) 0.094 (0.037)

R-squared 0.094 0.047

Notes: In this table, I regress Hirschman-Herfindahl (HHI) indices at differentgeographical levels on the consolidation treatment variable interacted with thepost-treatment dummy. The panel is defined at the market-year level. I controlfor market fixed effects and a linear time trend.

80

Table A3: Translog model in labor and capital

Panel (a): Output elasticities Labor CapitalEstimate: SE: Estimate: SE:

0.353 (0.226) 0.690 (0.140)

Panel (b): Markdowns and markup Markdown MarkupsEstimate: SE: Estimate: SE:

Mean 2.428 (1.28) 1.143 (0.230)

Panel (c): Consolidation effects log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment *1(year≥2003) 0.291 (0.061) -0.247 (0.061) 0.078 (0.083)

R-squared 0.799 0.751 0.866Observations 1,091 1,091 1,091

Notes: Panel (a) reports the output elasticities which are obtained from estimating the production function.Panels (b) and (c) report the estimated price semi-elasticities of input supply, both for tobacco leaf and labor. Theleft-hand side variable is the log province-level market share minus the log outside option market share. The right-hand side variable is the input price in levels. Manufacturing TFP from the Leontief model is used as instrument.Bootstrapped standard errors are in the parentheses.

81

Table A4: Using cost shares as output elasticities

Panel (a): Output elasticities Labor CapitalEstimate: SE: Estimate: SE:

0.427 ( ) 0.573 ( )

Panel (b): Markdowns and markup Markdown MarkupsEstimate: SE: Estimate: SE:

Mean 2.196 ( ) 1.25 ( )

Panel (c): Consolidation effects log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment *1(year≥2003) 0.268 (0.057) -0.222 (0.073) 0.081 (0.081)

R-squared 0.802 0.774 0.883Observations 1,091 1,091 1,091

Notes: This table reports the output elasticities, markdowns, markups and consolidation effects when assumingconstant returns to scale and when using the cost shares of labor and capital as output elasticities. Panel (a) reportsthe output elasticities, panel (b) the markdown and markup levels, and panel (c) the estimated consolidationtreatment effects at the province level.

82

Table A5: Overidentifying restrictions

log(Leaf market share) log(Leaf market share)Estimate: SE: Estimate: SE:

3.668 (0.862) 3.659 (1.793)Instruments:

Log(TFP) Yes YesConsolidation dummies No Yes

Sargan-Hansen test:χ2 4.332p-value 0.115

R-squared 0.053 0.147Observations 1,091 1,091

Notes: I estimate the semi-elasticity of leaf supply using either only manufacturing TFPas instrument for prices, or both the consolidation dummies and TFP as instruments. Ireport the Sargan-Hansen test statistic for overidentifying restrictions.

83

Table A6: Alternative input supply models

Panel (a): Log-on-logs log(Leaf market share) log(Labor market share)Estimate: SE: Estimate: SE:

log(Input price) 1.657 (0.207) 0.032 (0.187)

R-squared 0.723 0.8381st stage F-statistic 176 64Observations 1,091 1,091

Panel (b): Prefectural nests log(Leaf market share) log(Labor market share)Estimate: SE: Estimate: SE:

Input price 1.836 (0.198) 0.025 (0.024)

Nesting elasticity 0.621 (0.060) 0.849 (0.201)

R-squared 0.780 0.865Observations 1,091 1,091

Panel (c): Ownership type nests nests log(Leaf market share) log(Labor market share)Estimate: SE: Estimate: SE:

Input price 3.320 (1.590) -0.013 (0.117)

Nesting elasticity 0.545 (0.032) 1.253 (0.201)

R-squared 0.382 0.688Observations 1,091 1,091

Notes: In panel (a), I re-estimate the input supply models, but with log input prices on the right-handside, rather than input prices in levels. In panel (b), I define provinces as markets and prefectures assub-markets among which farmers can choose to be active in. In panel (c), I estimate the nested logitmodels with nests being defined as three ownership types: SOEs, private firms and collectives, andforeign firms.

84

Table A7: Markups and markdowns in nested logit models

Panel (a): Prefectural nestsMarkdown Markup

Estimate: SE: Estimate: SE:

Mean 3.049 (0.103) 0.960 (0.668)

log(Markdown) log(Markup)Estimate: SE: Estimate: SE:

Treatment * 1(year≥2003) 0.342 (0.060) -0.285 (0.055)

Panel (b): Ownership nestsMarkdown Markup

Estimate: SE: Estimate: SE:

Mean 2.852 (0.164) 1.052 (0.429)

log(Markdown) log(Markup)Estimate: SE: Estimate: SE:

Treatment * 1(year≥2003) 0.370 (0.059) -0.309 (0.057)

Notes: Panel (a) reports average markups, markdowns and treatment effectswhen using the geographically nested logit leaf supply model. Panel (b) doesthe same with the ownership nested logit model.

85

Table A8: Alternative treatment measures

Panel (a): Other treatment measures log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

i. Share firms under 100K

Province level 0.795 (0.144) -0.349 (0.154) 0.353 (0.203)Prefecture level 0.764 (0.140) - 0.411 (0.129) 0.154 (0.141)County level 1.120 (0.184 ) -0.579 (0.145) 0.069 (0.124)

ii. Employment share under 100K

Province level 0.270 (0.261 ) -0.353 (0.272) 0.482 (0.291)Prefecture level 0.719 (0.215 ) -0.465 (0.163) -0.156 (0.205)County level 1.035 (0.192) -0.559 (0.154) 0.002 (0.142)

Panel (b): Different fixed effects log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

ii. No firm fixed effects

Province level 0.403 (0.141) -0.267 (0.147) 0.059 (0.102)Prefecture level 0.261 (0.114) -0.038 (0.106) 0.123 (0.112)County level 0.729 (0.205) -0.332 (0.167) 0.210 (0.167)

iv. Adding year fixed effects

Province level 0.270 (0.059) -0.213 (0.056) -0.020 (0.068)Prefecture level 0.379 (0.070) -0.199 (0.066) 0.056 (0.067)County level 0.802 (0.144) -0.380 (0.113) 0.037 (0.102)

Panel (c): Different moments log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

v. Median

Province level 0.149 (0.022) -0.140 (0.051) -0.033 (0.052)Prefecture level 0.333 (0.081) -0.114 (0.065) 0.120 (0.063)County level 0.369 (0.078) -0.283 (0.142) 0.109 (0.099)

vi. Employment-weighted average

Province level 0.296 (0.064) -0.240 (0.068) -0.057 (0.106)Prefecture level 0.226 (0.086) -0.073 (0.083) 0.111 (0.075)County level 0.635 (0.131) -0.406 (0.121) 0.169 (0.101)

Notes:

86

Table A9: Announcement effects

Panel (a): Province level log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment * 1(year≥2001) 0.011 (0.083) -0.161 (0.064) -0.081 (0.065)

Treatment * 1(year≥2002) 0.164 (0.046) -0.008 (0.055) -0.065 (0.089)

Treatment * 1(year≥2003) 0.144 (0.057) -0.128 (0.060) 0.067 (0.106)

Panel (b): Prefecture level log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment * 1(year≥2001) 0.143 (0.081) -0.214 (0.075) -0.096 (0.057)

Treatment * 1(year≥2002) 0.123 (0.075) 0.031 (0.070) -0.058 (0.075)

Treatment * 1(year≥2003) 0.224 (0.082) -0.124 (0.083) 0.141 (0.089)

Panel (c): County level log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment * 1(year≥2001) 0.057 (0.085) -0.129 (0.082) -0.125 (0.060)

Treatment * 1(year≥2002) 0.268 (0.112) -0.150 (0.100) -0.029 (0.097)

Treatment * 1(year≥2003) 0.529 (0.168) -0.185 (0.137) 0.111 (0.120)

Notes: I re-estimate the difference-in-differences model from equation (1). In contrast with the mainspecification, I define the start of the treatment effect to take place in 2000, 2001 and 2002. Province-,prefecture- and county-level market definitions are used in panels (a), (b) and (c). The specific sizethresholds were announced in 2002.

87

Table A10: Narrowing the time frame

Panel (a): 1999-2005 log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment * 1(year≥2003) 0.271 (0.071) -0.239 (0.068) 0.094 (0.093)

1(year≥2003) -0.274 (0.074) 0.252 (0.073) -0.166 (0.100

Observations 984 984 984R-squared 0.806 0.787 0.869

Panel (b): 1999-2004 log(Markdown) log(Markup) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment * 1(year≥2003) 0.199 (0.084) -0.160 (0.081) 0.063 (0.116)

1(year≥2003) -0.214 (0.083) 0.186 (0.081) -0.150 (0.115)

Observations 924 924 924R-squared 0.817 0.796 0.883

Notes: In panels (a) and (b), I re-estimate the consolidation treatment effects using shorter time framesfor the post-treatment period. I reduce this treatment period from three to two and one years, respectively.

88

Table A11: Sample size robustness

Panel (a): Dropping firms Q < 100K log(Markdown) log(Markup) log(TFP)

Treatment * 1(year≥2003) 0.223 (0.065) -0.243 (0.065) -0.006 (0.081)

1(year≥2003) -0.165 (0.073) 0.170 (0.073) -0.051 (0.093)

Observations 626 626 626R-squared 0.767 0.817 0.668

Panel (b): Using ACF sample size Labor coefficient Capital coefficient Scale economies

Treatment * 1(year≥2003) 0.365 (0.155) 0.804 (0.148) 1.168 ()

Observations 825R-squared 0.903

Notes: In panel (a), I drop the surviving firms under the exit threshold and re-estimate the treatment effectsmodel on this sample. In panel (b), I re-estimate the production function using the IV approach but reduce thesample to the ACF estimation sample.

89

Table A12: Exit and mergers

Panel (a) Province-level: log(Markdown) log(Markup) log(TFP)

Exit treatment * 1(year≥2003) 0.320 (0.113) -0.261 (0.103) 0.034 (0.128)

Merger treatment* 1(year≥2003) -0.043 (0.117) 0.017 (0.113) 0.081 (0.141)

Observations 1,091 1,091 1,091R-squared 0.801 0.756 0.894

Panel (b) Prefecture-level: log(Markdown) log(Markup) log(TFP)

Exit treatment * 1(year≥2003) 0.381 (0.103) -0.197 (0.087) 0.133 (0.094)

Merger treatment* 1(year≥2003) 0.044 (0.107) -0.077 (0.090) -0.029 (0.094)

Observations 1,091 1,091 1,091R-squared 0.880 0.844 0.895

Panel (c) County-level: log(Markdown) log(Markup) log(TFP)

Exit treatment * 1(year≥2003) 0.721 (0.144) -0.353 (0.120) 0.100 (0.108)

Merger treatment* 1(year≥2003) 0.192 (0.104) -0.151 (0.094) -0.010 (0.079)

Observations 1,091 1,091 1,091R-squared 0.879 0.868 0.894

Notes: In this table, I re-estimate the consolidation treatment model with two types of treatmenteffects. The ‘exit treatment’ is the same consolidation dummy as used throughout the main text.It indicates the presence of at least one firm below the exit threshold of 100,000 cases produced in2002. The ‘merger treatment’ indicates the presence of at least one firm below the merger thresholdof 300,000 cases per year. Panels (a)-(c) re-estimate the treatment effects for all three outcomes ofinterest at the province, prefecture and county-level respectively.

90

Table A13: Consolidation treatment effects, controlling for the capital-labor ratio

Panel (a): Province level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.299 (0.063) -0.256 (0.064) 0.076 (0.083)

log(Capital/labor) 0.007 (0.028) 0.044 (0.050) -0.199 (0.059)

R-squared 0.808 0.762 0.870Observations 1,091 1,091 1,091

Panel (b): Prefecture level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.407 (0.071) -0.246 (0.069) 0.101 (0.073)

log(Capital/labor) 0.081 (0.058) -0.029 (0.056) -0.197 (0.059)

R-squared 0.880 0.847 0.870Observations 1,091 1,091 1,091

Panel (c): County level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

log(Capital/labor) 0.002 (0.063) 0.036 (0.060) -0.198 (0.059)

Treatment * 1(year≥2003) 0.788 (0.134) -0.412 (0.111) 0.082 (0.102)

R-squared 0.880 0.870 0.870Observations 1,091 1,091 1,091

Notes: In this table, I re-estimate the difference-in-differences model using various input marketdefinitions, while controlling for capital intensity differences across firms.

91

Table A14: Comparison to prior monopsony studies

Paper Country Industry Input Input price / MRP

Goolsbee and Syverson (2019) USA Universities Professors 0.83Ransom and Sims (2010) USA Schools Teachers 0.82 (M) - 0.75 (F)Hirsch, Schank, and Schnabel (2010) Germany All Employees 0.78Oaxaca and Ransom (2010) USA Grocery stores Clerks 0.74 (M) - 0.70 (F)This paper China Tobacco Farmers 0.38

Note: I report the average price of an input over its marginal revenue product, which I calculate based on thereported labor supply elasticity using equation (9).‘M’ stands for male and ‘F’ for female workers.

92

Table A15: Parametrization of Monte Carlo simulations

Parameter Description Value

βl Output elasticity of labor 0.4βk Output elasticity of capital 0.6ωjt TFP ∼ exp(N(0, 0.5))ζjt TFP shock ∼ exp(N(0, 0.5))ρ TFP serial correlation 0.7σa TFP cross-sectional standard deviation 0.5Mm Input market size [0.5,1.5]ξjt Working conditions ∼ U [0, 3]αξ Working condition valuation 0.5αw Wage valuation 2.5γw Input price valuation by supplier (DGP 1) 0γw Input-price valuation by supplier (DGP 2) ∼ U [−0.375, 0.375]K0 Initial capital stock ∼ exp(N(0, 0.25))κ Investment cost coefficient 1δ Depreciation 0.1

93

Table A16: Monte Carlo simulations

Panel (a): Output elasticities OLS ACF ACF bis∗

– standard logit supply model Estimate: SE: Estimate: SE: Estimate: SE:

Labor (truth = 0.4) 0.544 (0.028) 0.582 (0.070) 0.401 (0.113)

Capital (truth = 0.6) 0.555 (0.025) 0.543 (0.030) 0.600 (0.043)

Panel (b): Output elasticities OLS ACF bis∗

– mixed logit supply model Estimate: SE: Estimate: SE:

Labor (truth = 0.4) 0.518 (0.029) 0.489 (0.127)

Capital (truth = 0.6) 0.562 (0.024) 0.564 (0.041)

Panel (c): Input supply elasticity OLS IVEstimate: SE: Estimate: SE:

Labor (truth = 2.5) -0.159 (0.083) 2.249 (0.641)

First-stage F-stat – – 97.949 (76.044)

Notes: *ACF with input market shares and input prices in first stage. Instrument: productivity residuals.Controls: market dummies. Simulated using 50 iterations.

94

Table A17: Elasticity of substitution

log(Materials/Labor) Log(Capital/Labor)Estimate: SE: Estimate: SE:

Elasticity of substitution 0.146 (0.281) 0.909 (0.222)

First-stage F-stat 64.58 64.58

Observations 1,091 1,091R-squared 0.283 0.478

Notes: Controls: ownership type, product type, year dummies, export dummy,export revenue share. IVs: avg. export share of revenue of competitors, exportparticipation of competitors.

95

Table A18: Ownership consolidation and labor-augmenting productivity

Panel (a): Province-level markets log(LAP∗)Estimate: SE:

Treatment * 1(year≥2003) 0.228 (0.089)

Observations 1,091R-squared 0.662

Panel (b): Prefecture-level markets log(LAP∗)Estimate: SE:

Treatment * 1(year≥2003) 0.141 (0.089)

Observations 1,091R-squared 0.661

Panel (c): County-level markets log(LAP∗)Estimate: SE:

Treatment * 1(year≥2003) 0.042 (0.120)

Observations 1,091R-squared 0.661

Notes: ∗Labor-augmenting productivity. Dependent variables inlogs. Markup re-calculated with labor-augmenting productivitytaken into account.

96

Table A19: Product characteristics

Panel (a): Cigarette content log(Leaf weight) log(Filter density) log(Rod density)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment 0.000107 (0.00685) 0.00571 (0.0101) 0.0142 (0.0181)

Observations 353 353 353R-squared 0.804 0.702 0.550

Panel (b): Cigarette design log(Paper permeability) log(Ventilation)Estimate: SE: Estimate: SE: Estimate: SE:

Treatment 0.0331 (0.0732) -0.891 (0.491)

Observations 353 353R-squared 0.586 0.848

Panel (c): Correlations log(Markup) log(Markdown) log(TFP)Estimate: SE: Estimate: SE: Estimate: SE:

Ventilation 0.0149 (0.0281) 0.0261 (0.0303) -0.00883 (0.0573)

Rod density 4.685 (6.984) -0.253 (7.524) 1.224 (14.24)

Filter density 22.46 (6.346) -8.990 (6.836) -9.010 (12.93)

Leaf weight 11.66 (3.651) -5.984 (3.933) -9.399 (7.441)

Paper permeability 1.981 (1.355) -1.391 (1.459) -0.786 (2.761)

Observations 137 137 137R-squared 0.687 0.700 0.488

Notes: Panel (a) compares the cigarette contents between the treatment and control groups. Panel (b) does the samefor two cigarette design features. Panel (c) reports the correlations between markups, markdowns, productivity andcigarette characteristics. Province dummies are controlled for.

97

Table A20: Consolidation and exporting

Export dummy Log(Exports/Revenue)Estimate: SE: Estimate: SE:

Treatment * 1(year≥2003) -0.048 (0.043) -0.011 (0.009 )

Treatment -0.030 (0.042) -0.006 (0.010)

1(year≥2003) 0.076 (0.025) 0.008 (0.003)

Observations 2,638 2,486R-squared 0.040 0.0111(year≥2003) 0.076 (0.025) 0.008 (0.003)

Note: Controls include a time trend, product type, ownership type and firm fixedeffects.

98

Table A21: Consolidation effects, alternative market definitions

Panel (a): Province level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.299 (0.063) -0.255 (0.064) 0.074 (0.082)

R-squared 0.801 0.788 0.862Observations 1,091 1,091 1,091

Panel (b): Prefecture level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.405 (0.071) -0.245 (0.069) 0.107 (0.074)

R-squared 0.879 0.847 0.867Observations 1,091 1,091 1,091

Panel (c): County level log(Markdown) log(Markup) log(Productivity)Estimate S.E. Estimate S.E. Estimate S.E.

Treatment * 1(year≥2003) 0.788 (0.134) -0.413 (0.111) 0.084 (0.103)

R-squared 0.718 0.763 0.862Observations 1,091 1,091 1,091

Notes: Estimates of treatment effect in difference-in-differences regression reported. Same specifi-cation as before. Dependent variable in logs. Treatment and control groups calculated by definininginput markets at different geographical levels.

99

Table A22: Market power and ownership

Panel (a): Buying power log(Markdown) log(Markdown)Estimate: SE: Estimate: SE:

Log(% Equity state-owned) 0.135 0.0240 0.0168 (0.0618)

Log(% Equity owned by legal person) 0.246 (0.079) 0.101 (0.077)

Observations 1,091 1,091R-squared 0.061 0.877Firm FE No Yes

Panel (b): Selling power log(Markdown) log(Markdown)Estimate: SE: Estimate: SE:

Log(% Equity state-owned) 0.310 (0.0240) 0.0135 (0.0618)

Log(% Equity owned by legal person) 0.246 (0.0413) 0.101 (0.125)

Observations 1,091 1,091R-squared 0.044 0.743Firm FE No Yes

Notes: Omitted ownership categories are ’private person’, ’foreign’ and ’collective’.

100